ohgeebtw/geezuz-qwen3.8-27b-dual-gpu is a repository that runs Qwen3.8-27B GGUF weights on two GPUs with a CUDA megakernel. A CUDA megakernel is a way of bundling multiple GPU operations into one kernel. The latest release is v0.3.2. The repository uses tensor parallelism, meaning the model is split across several GPUs and computed together. It is a fork that extends the single-GPU megakernel of open-jet to two GPUs over PCIe P2P.
The structure of this repository assumes multiple requests. It decodes up to 8 requests together in one cycle, and several requests share a single KV cache pool and prefix cache. The KV cache holds the keys and values of previous tokens needed for attention computation, and the prefix cache is a store that reuses a repeated leading part instead of computing it again. On-device MTP speculative decoding produces candidate tokens first, verifies them, and then confirms several tokens at once. It supports Q4_K_M, Q5_K_M, Q6_K, Q8_0 GGUF weight formats, and the repository includes a command-line runner and an OpenAI-compatible server. The README reports an average 1.64× speed over llama.cpp together with top-token match rate, KL and perplexity comparisons. The license is AGPL-3.0 and the primary language is Cuda.
However, the changes in v0.3.2 are not included in the source material, so what differs in this version cannot be confirmed. The 1.64× figure in the README was measured by the author in a specific GPU configuration, and the published performance table does not generalize to other environments. Until an external reproduction is confirmed, it is closer to a performance claim than an established speed advantage. Given that this is a 7-star repository created in October 2026, it is still too early to conclude that it is faster than other implementations.