ONNX Runtime GenAI model-builder quantization paths (int4, fp8, ModelOpt/compressed-tensors ingest)

PROXY ENUM VALUE - DECLARED 2026-10-04: this is an INGEST/EXPORT PATH record (MatMulNBits int4 plus ModelOpt and compressed-tensors ingest), filed under 'int8'. The int4 part is covered precisely by the separate ort-matmulnbits-int4 record at group_size 128, so 'int8' is wrong even as an approximation here - the record's own bits field reads 'int4 (MatMulNBits weight-only, block size selectable)'. A query for INT8 returns a record whose primary path is INT4. REQUESTED ENUM ADDITION: 'ortgenai-model-builder'. RECORDED AS THE QUANT PATHS, NOT AS AN ENGINE: an engine record for onnxruntime-genai already exists, so this record covers only what formats it accepts and emits. INPUT PATHS verified from the builder README: (1) a dense HF checkpoint with -p precision; (2) an ALREADY int4-quantized PyTorch checkpoint loadable via AutoGPTQ or AutoAWQ with -p int4; (3) an existing float16/float32 GGUF on disk, which the builder consumes directly - so GGUF is an INPUT to ORT-GenAI, not only an output of llama.cpp. AWQ and GPTQ usage is redirected here from the ORT quantization page. OUTPUT PATHS: the builder can export a ModelOpt or compressed-tensors NVFP4/FP8 checkpoint, which is the interoperability bridge from Microsoft's stack to Red Hat's - ORT-GenAI can emit the compressed-tensors format and llmcompressor can emit ModelOpt/compressed-tensors NVFP4. PER-TENSOR-ROLE OPTIONS: qmoe_block_size and qmoe_weights_prepacked, matmulnbits_block_size, accuracy_level, is_symmetric, op_types_to_quantize, nodes_to_exclude, algo_config, int8_bit_placement, use_qdq_pattern, and use_8_bits_quantization_in_qmoe / use_fp4_quantization_in_qmoe - the last two are notable because they make bit-width a PER-EXPERT-CHOICE inside a single MoE graph, which is the ONNX-side equivalent of the mixed-precision MoE literature. PROVENANCE AND STATUS: builder_config_version 2 is explicitly labelled an EXPERIMENTAL implementation of the shared configuration design; the README also lists 'Olive integration, target checkpoint conversion policy, and INT8 embedding export remain pending'. So do not assume Olive and the ORT-GenAI builder currently compose.