juraj-jumic/dspark-local-bench is a reproducible benchmark repository that compares what difference llama.cpp's speculative decoding makes in a single-user environment on consumer graphics cards. The repository description targets measurements at batch-size 1, that is, the condition of handling one user at a time. The README provides the run commands and explains that results from other GPUs can be added as well. The focus of this project is comparison on a personal GPU, not server performance when many users connect at the same time.
The README presents a table of Vulkan and ROCm results using Qwen3-8B Q8_0 on a Radeon RX 7900 GRE. The compared items are DSpark, DFlash, and a small draft model. It examines results by draft token count and acceptance threshold separately from prompt processing cost, and also covers differences in the output. So instead of simply listing generation speed, it is material that compares together the aspects that vary depending on the speculative decoding method and settings. The repository language is Python and the license is MIT.
However, the results stated in the README are limited to a single GPU and a specific prompt configuration. From the presented table alone, one cannot conclude that the same results would appear on other graphics cards or under other input conditions. The current material is better read as a reference for looking at the differences between methods under that Radeon RX 7900 GRE configuration.