nvidia

/

llama-nemotron-embed-vl-1b-v2

like 59

Follow

NVIDIA 55.3k

• Add vLLM usage instructions to README covering online serving (/v1/embeddings API with chat template) and offline/in-process (llm.embed()) usage
• Examples for all supported modalities: text query, text document, image document, and image+text document
• Verified outputs match transformers reference (>0.9999 cosine similarity across all modalities)

Test plan

• Offline vLLM embedding produces 2048-dim vectors for all four modalities
• Online vLLM server returns 200 on /v1/embeddings for all four modalities
• Cosine similarity between transformers and vLLM (offline + online) embeddings >0.9999 for all embedding types