KTransformers as the framework for heterogeneous LLM inference and fine-tune optimizations is out today with its v0.7 feature release.

With KTransformers 0.7 there is now full AVX-512 support for LoRA fine-tuning without depending upon Advanced Matrix Extensions (AMX) also being present. This AVX-512-only without AMX benefits AMD EPYC Zen 4 / Zen 5 / Zen 6 servers with excellent AVX-512 support while lacking AMX and also older Intel Xeon processors with AVX-512 prior to the introduction of AMX with Sapphire Rapids.

The merge request noted the testing on AMD hardware and the foxus on AVX-512 without AMX platforms. The KTransformers runtime will automatically select the proper CPU implementation and in turn allowing MoE expert training to happen on a wider range of large-memory servers.

KTransformers 0.7 also adds VLM fine-tuning support, native FP8 LoRA support, improved DeepSeek V4 deployment, and CPU activation reuse.

More details on KTransformers 0.7 for those using it for LLM inference optimizations and fine-tuning can find all the details via the release announcement on GitHub.