Infinity v3.0.0-alpha.5

Alpha release of Infinity v3.

Changes since alpha.4:

• Add image inputs to OneJev block-attention checkpoints, preserving mRoPE coordinates and expanded prefix boundaries (#719).
• Reuse shared visual prefixes, KV/GDN state and image features across prefix and suffix passes while retaining packed execution and optimized kernels. Native block-image decision calls now require explicit mRoPE coordinates (#719).
• Add LibTorch Mistral3ForSequenceClassification reranker support, including Nemotron 3.5 Rerank (#717).
• Add a Truss deployment example for Laya on two L4 GPUs (#718).

Validation for OneJev images includes native CPU/H100 tests, independent full-sequence Transformers comparisons, cross-pass vision-reuse profiling, and HTTP image/token-billing checks on a local trained 9B checkpoint. Prefix reuse is local to a native batch; block-mode audio, radix folding and persistent cross-request caching remain unsupported.

Images:

docker pull ghcr.io/michaelfeil/infinity_v3:nvidia-3.0.0-alpha.5
docker pull ghcr.io/michaelfeil/infinity_v3:cpu-3.0.0-alpha.5

Use the NVIDIA LibTorch image for OneJev multimodal support. The CPU image uses Candle. Images publish asynchronously through the tag-triggered Infinity v3 Docker workflow; wait for both builds and image-start checks to pass before pulling the new tags.

LibTorch guide · Laya Truss example · Licensing and component notices · Full changelog