dummy-cyclic-llava-ov

This is not a real vision-language model. It is a synthetic, hand-crafted LLaVA whose text backbone was built to deterministically emit one fixed sentence followed by a valid end-of-sequence token, regardless of the input prompt or the input image. It exists purely as a fast, tiny, deterministic drop-in for testing OVMS / OpenVINO GenAI integration code. Do not use it for anything that expects real vision or language understanding.

What it outputs

Regardless of the prompt and regardless of the image passed in, this model deterministically generates the exact same sentence as its LLM sibling model (dummy-cyclic-gpt2-ov):

OpenVINO is an open-source toolkit created by Intel to speed up and run AI models efficiently.

With normal generation settings (the default - ignore_eos unset/False), the model emits this sentence once and then a valid end-of-sequence token, so generation stops there. It only repeats forever if the caller explicitly sets ignore_eos=True with a large enough max_new_tokens budget - see dummy-cyclic-gpt2-ov's model card for exactly why.

Image tokens are injected at early sequence positions by the vision encoder/projector, but because the text backbone's attention is zeroed (see below), they never influence the hidden state of the last text token - so the generated output is completely image-agnostic. Text-side behavior (where generation starts/resumes depending on the prompt, and whether it stops or loops) is identical to dummy-cyclic-gpt2-ov.

How it behaves

• Architecture: LlavaForConditionalGeneration - a LLaMA text backbone (1 layer, 1 attention head, hidden size 32, RMSNorm) plus a small CLIP vision encoder and multimodal projector.
• The text backbone uses the identical deterministic mechanism as dummy-cyclic-gpt2-ov: token embeddings map to unique orthogonal directions, attention/MLP weights are zeroed (making every block a pure residual/identity), and lm_head deterministically predicts the next token of the fixed sentence above with a dominant (~100) logit - ending, on the last token, in a valid end-of-sequence prediction rather than a wraparound.
• Unlike GPT-2, LLaMA has no learned position-embedding table - it uses RoPE instead. Since attention is zeroed anyway, RoPE has no effect on the output, so generation is independent of position exactly like the LLM sibling model.
• The CLIP vision encoder and multimodal projector keep their random, untrained default weights. They still run and produce valid image embeddings, but those embeddings never affect the generated text.

Supported OVMS pipeline types

VLM, VLM_CB (image-text-to-text task, with or without continuous batching).

License

Apache 2.0. No pretrained weights are used in the text backbone - those weights are hand-crafted from scratch. The CLIP vision encoder and multimodal projector keep their randomly-initialized (untrained) default weights, so no upstream pretrained-model license applies to those either.