jchen-318/gaudi2-single-card-strata is an inference engine repository that runs Qwen3.8-Flash-Next natively on a single Intel Gaudi2. It is a project built on Strata, and the creation date recorded in FACTS is October 21, 2026. There is no release information. It compiles the model layer into a SynapseAI graph and runs it on Gaudi2's MME and TPC, and it includes around a hundred hand-written TPC kernels. The code is C++ and the license is MIT.
The purpose of this repository is to run Qwen3.8-Flash-Next inference on a single Gaudi2 card. The target model has a 512-expert MoE architecture and a 256k context. The README describes DFlash2 speculative decoding and an OpenAI- and Anthropic-compatible API. It also provides a Docker configuration, a CMake build, unit tests, and hardware test procedures. The performance listed in the README is decode ~155 tok/s and prefill ~3.5k tok/s.
The performance figures were measured by the repository's author on specific hardware under single-request conditions, so they are not comparative data showing performance in other environments. According to the README, only one model architecture is supported, and multiple requests are handled one at a time. The scope this record shows therefore extends as far as an implementation that runs that model on one Gaudi2.