1115714829/llama.cpp-v100 is a project that modifies and optimizes llama.cpp's compute kernels and runtime for NVIDIA Tesla V100 (SM70). The latest release confirmed in the provided FACTS is v100-2.0.1.1. However, the specific changes in this version are not given. The repository description lists multi-GPU execution and long-context handling, multimodal processing, concurrent serving, and a built-in web UI as its supported scope.
This fork focuses on providing multiple inference paths on V100 without changing public GGUF weights or quantization formats. According to the description, Qwen3.8-Flash-Next records a 200K-context prefill of 3402 tok/s on a 6× V100-16GB configuration, and it also runs on a 2-GPU configuration. Qwen3.8-27B is introduced as supporting up to 220 tok/s decoding with DFlash2 speculative decoding. Speculative decoding generates candidate next tokens first and then verifies them. The repository states support for 2–6 GPUs and 262K–512K contexts.
These figures come from the V100 configuration stated in the repository description and do not show performance on other GPUs. Also, the latest release number is confirmed, but the changelog for v100-2.0.1.1 is not in the given material. Therefore, it is difficult to judge the extent of the improvements in that version or the performance gap against other GPUs from this material alone.