michaelfeil/infinity v3.0.0-alpha.9

Infinity v3.0.0-alpha.9

on GitHub

Changes since alpha.8:

• Enable opt-in dynamic FP8 for dense Qwen3.5 models, including qwen3_5 and qwen3_5_text. Use CUDA BF16 with --enable-fp8-dynamic; INFINITY_TORCH_FP8_SCOPE=mlp selects MLP-only quantization (#735).
• Accept criteria: {} for OneJev yes/no (noul) questions, using the default yes/no descriptions (#734).
• Honor modeling_overrides decision attention metadata and queue multimodal downloads instead of immediately rejecting requests when download slots are occupied (#733).

CPU and NVIDIA image builds run asynchronously; the tags become available after publication:

docker pull ghcr.io/michaelfeil/infinity_v3:nvidia-3.0.0-alpha.9
docker pull ghcr.io/michaelfeil/infinity_v3:cpu-3.0.0-alpha.9

FP8 changes numerical outputs and remains opt-in. Local H100 tests passed; PR #735 was merged with admin override while remote CI and review were still pending.

Image builds · Licensing · Full changelog