embeddinggemma-2-onnx-fp16

FP16 mirror of the EmbeddingGemma 2 text tower for CUDA serving with embroider.

• Source: onnx-community/embeddinggemma-2-ONNX @ daa72c5 (onnx/model_fp16.onnx + data), kept at onnx/model_fp16.onnx layout (same convention as in-repo nano model_fp16.onnx; the graph's external-data path names this file).
• Base weights: google/embeddinggemma-2 (Apache-2.0). Text tower only (270M); vision/audio encoders out of scope.
• Contract: Gemma prompts (task: search result | query: / title: none | text:), mean pooling + projection + L2 in-graph (sentence_embedding), MRL 128/256/512/768, 8K context.
• Parity (CPU, ORT 1.29): worst 2.1e-4 vs fp32, cosine ~1.0 on 4 query/document probes. FP16 is CUDA-only by design (CPU emulated, slow); serve fp32 on CPU.
• Files: onnx/model_fp16.onnx + onnx/model_fp16.onnx_data, tokenizer.json, tokenizer_config.json, config.json, config_sentence_transformers.json, manifest.json (SHA-256).