{ "primary_runtime": { "kind": "direct ONNX Runtime CUDA InferenceSession", "provider": "CUDAExecutionProvider (NVIDIA H200)", "onnxruntime": "1.27.0 (onnxruntime-gpu)", "status": "WORKS", "evidence": ["evidence/l4_parity.json", "evidence/l5_generation.json", "evidence/l5_assisted_detail.json", "evidence/l5_drafter_rootcause.json"] }, "canonical_metadata_schema": { "schema": "onnx-genai v1 (schema/inference_metadata.schema.json on PR #1716)", "target": "SHIPPED + VALIDATED: target/package/inference_metadata.yaml is a faithful v1 decoder workflow (real past_key_values.N/present.N ports, full owners 4/9/14, 12 sliding owners) that passes the #1716 JSON Schema (draft 2020-12).", "assistant": "canonical v1 speculative shape exists (onnx-genai #1716 example 24) but a faithful real-artifact YAML is blocked by 3 deltas (see assistant METADATA_STATUS.md): StatePortAlias.output required vs cacheless drafter with no present_* KV; two-repo layout vs combined speculative package; single shared_kv. vs per-owner target ports.", "note": "read_only aliases (#1700) + chained proposal_execution (#1696) are integrated in #1716 but NOT on onnx-genai main; validate against #1716." }, "onnxruntime_genai_genai_config": { "status": "NOT FAITHFUL / NOT RUNNABLE here", "reasons": [ "onnxruntime_genai 0.14.1 in this env is the CPU build; libonnxruntime-genai-cuda.so is absent, so CUDA generation cannot initialize.", "genai_config.json exposes a single flat decoder.head_size (emitted as 256); the target has HETEROGENEOUS head_dim (sliding/local=256, full/global=512 at layers 4,9,14), which genai_config cannot represent. The 3 full-attention layers cannot be driven faithfully through a uniform head_size, so the hybrid target is not faithfully runnable via the released genai_config path." ] }, "onnx_genai_engine_inference_metadata": { "status": "DEFERRED (engine not installed here)", "reasons": [ "The onnx-genai engine that consumes inference_metadata.yaml is present only as a Rust source workspace (/home/justinchu/onnx-genai-p2-fmt); no released binary matched to the #1716 schema is installed here.", "The shipped target v1 YAML was validated with a standalone JSON-Schema (draft 2020-12) check against the #1716 schema, not by executing the engine." ] } }