Changes since alpha.6:
- Add InfinityFA4 native attention and DeBERTa fusion, with bundles for SM80, SM86, SM89, SM90, SM100, and SM120 (#726).
- Separate LibTorch encoder classifier readout (#727).
- Add LibTorch Llama and Qwen token classification (#728).
- Reduce the LibTorch runtime image by selecting required shared libraries and sharing identical FA4 runtimes across architecture bundles (#729).
CPU and NVIDIA image builds are running asynchronously. Check the workflow for publication and startup-check results:
docker pull ghcr.io/michaelfeil/infinity_v3:nvidia-3.0.0-alpha.7
docker pull ghcr.io/michaelfeil/infinity_v3:cpu-3.0.0-alpha.7