Shack870/qwen-air-qpu-mcp-lab
收藏资源简介:
该数据集仓库镜像了Qwen Air QPU/MCP Lab的公共工件:这是一个在2017年Intel MacBook Air(配备8GB RAM)上进行的CPU-only Qwen3 MoE推理优化项目。关键结果是在16,384上下文长度下,使用同步的Codex驱动、IBM Quantum-informed autoresearch循环,将Qwen3-30B-A3B-Instruct-2507-GGUF的生成令牌/秒从约0.09提升到14.03。QPU并未运行Qwen;IBM Quantum在研究循环内采样紧凑的QUBO候选空间,而MacBook通过本地的llama.cpp推理对每个候选进行评估。
This dataset repository mirrors the public artifacts for the Qwen Air QPU/MCP Lab: a CPU-only Qwen3 MoE inference optimization project on a 2017 Intel MacBook Air with 8GB RAM. The headline result is a quality-gated improvement from about 0.09 to 14.03 generation tokens/sec on Qwen3-30B-A3B-Instruct-2507-GGUF at 16,384 context, using a synchronized Codex-driven, IBM Quantum-informed autoresearch loop. The QPU did not run Qwen. IBM Quantum sampled compact QUBO candidate spaces inside the research loop; the MacBook judged each candidate through real local llama.cpp inference.




