EmbeddingGemma 2 Jetson Runtime
Runtime 0.1.3 for NVIDIA Jetson AGX Thor and AGX Orin. These archives contain the inference service and its dependencies; Google model weights and tokenizer files are downloaded separately and mounted read-only at /model.
This release fixes the required created field in /v1/models, adds JSON and SSE responses to /v1/chat/completions for connectivity probes, and accepts embedding requests of up to 2,048 inputs while processing bounded GPU batches. The chat endpoint describes the embedding service; it does not generate conversational answers. Each individual input retains the 8,192-token limit.
| Device | Archive | Bytes | SHA256 |
|---|---|---|---|
| Thor | embeddinggemma-2-thor-0.1.3.tar |
2,414,639,104 | 04e841144e4fea1fe0471598ea5804e2fa7cd749a3e3d9badfc411754b429c85 |
| Orin | embeddinggemma-2-orin-0.1.3.tar |
2,155,264,512 | e6630f9c731eb0068c1c9e47bad85d876be87cb4b61ed55003a351f246ae3f55 |
The Hugging Face and ModelScope mirrors publish identical archives. This release distributes downloads through these two hubs only; the installer does not use a Registry CDN fallback. runtime-manifests.json records the image identities and archive checksums; SHA256SUMS provides the two archive checksums. Older versioned archives remain available.
Both device builds passed the same 18-case GPU compatibility suite, including large batches, output ordering, vector dimensions, float/base64 encodings, text/image/audio/video inputs, JSON/SSE probes, and request boundaries. Five representative text and media embeddings matched the previous runtime with cosine similarity above 0.99999999999999. Model weights, tokenizer files, and their pinned Google source revisions are unchanged.
Use the EmbeddingGemma 2 AI Pod LPK 0.1.7 to select the download source and install the matching device archive. The installer verifies the archive size and SHA256 before importing the image, mounts the model files read-only, and starts offline inference. Runtime archives and model weights are external downloads and are not bundled in the LPK.