Synthetic LLaMA-style OpenVINO model from the NPUW model builder for NPU/NPUW testing. Random weights (untrained — output is gibberish by design). Projection MatMuls named optimum-style (<layer>/MatMul) so GenAI dynamic-LoRA works.
Geometry: 8 layers, hidden 2048, 16 heads, head_dim 128, inter 5632, vocab 32000, fp16, kv-cache on.
1hf download dylanneve1/npuw-synthetic-baseline --local-dir npuw-synthetic-baseline
2python openvino.genai/tools/llm_bench/benchmark.py --genai -m npuw-synthetic-baseline -d NPU -ic 16 --disable_prompt_permutation
