Local ONNX models with multi-shard external weights fail at ORT init; Harrier fp16 exports hit a CUDA GroupQueryAttention kernel gap

[lemon07r]

Found during the 2026-08-21 default-model screening (docs/benchmarks.md, "Default-Model Screening").

1. Multi-shard external weights unsupported. Vera's local-model config downloads a single ONNX external-data file (onnx_data_file). Models whose weights exceed one shard (convention: model.onnx_data, model.onnx_data_1, ...) fail at session init:

Exception during initialization: filesystem error: cannot get file size: No such file or directory [...]/onnx/model.onnx_data_1

Repro: lane {backend: custom-onnx, repo: onnx-community/harrier-oss-v1-0.6b-ONNX, onnx_file: onnx/model.onnx, onnx_data_file: onnx/model.onnx_data}. Workaround used for the screening: manual pinned download into a local dir lane.

2. fp16 GQA exports rejected by the CUDA EP. Harrier's model_fp16.onnx uses the fused GroupQueryAttention op with an attention bias, which the ORT CUDA kernel rejects:

Non-zero status code returned while running GroupQueryAttention node. Name:'/model/layers.0/attn/GroupQueryAttention' Status Message: attention_bias is not supported in GroupQueryAttention cuda kernel.

fp32 runs fine (the fusion is fp16-specific, so fp32 falls back to the unfused path). Today this only bites users who point VERA_LOCAL_EMBEDDING_ONNX_FILE at such an export; Vera's built-in GPU swap (quantized -> fp16) applies only to the jina preset, which exports a compatible graph.