--- license: apache-2.0 tags: [onnx, onnxruntime, onnx-genai, inference-metadata, peft, lora] --- # onnx-genai-example-qwen2-5-1-5b-lora-selection Public real-weight ONNX package from [`Qwen/Qwen2.5-1.5B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/tree/989aa7980e4cf806f80c7fef2b1adb7bc71aa306) at `989aa7980e4cf806f80c7fef2b1adb7bc71aa306`, with matching PEFT adapter [`bharati2324/Qwen2.5-1.5B-Instruct-Code-LoRA-r16`](https://huggingface.co/bharati2324/Qwen2.5-1.5B-Instruct-Code-LoRA-r16/tree/57a4a23b934ea6c3f25615e13a6979d55c48fd68) at `57a4a23b934ea6c3f25615e13a6979d55c48fd68`. Both sources are Apache-2.0. Includes actual fp16 decoder and adapter weights, canonical metadata, tokenizer, policies, source provenance, request/output, graph report, and H200 CUDA timings. The real probe used logical rows at scales 0, 0.5, and 1.0; adapter rows changed final logits by 25.72–25.79 and generated different output from the base row. Installed ORT exposes whole-run adapter activation, so heterogeneous logical rows were executed independently and this limitation is explicit in `output.json`. ## Download ```bash hf download justinchuby/onnx-genai-example-qwen2-5-1-5b-lora-selection --repo-type model --local-dir ./qwen2.5-1.5b-lora-selection ``` ## Exact runtime probe ```bash cd qwen2.5-1.5b-lora-selection python3 evidence/probe_lora_onnx.py cat evidence/lora_runtime_probe.json ``` Requires CUDA-capable `onnxruntime-gpu`, `onnx-ir`, `transformers`, `safetensors`, `torch`, and `numpy`; exact successful versions are in `output.json`. ## Annotated inference metadata Review [`inference_metadata.annotated.yaml`](./inference_metadata.annotated.yaml) for inline explanations of this package's workflow, tensor/state/cache contracts, and fail-closed omissions. [`inference_metadata.yaml`](./inference_metadata.yaml) remains the canonical machine-authored contract; automated validation confirms both files parse to the same metadata object.