Developer runs Qwen3-27B on dual RTX 3090s without NVLink, hits 210 tok/s after week of fixes

A developer spent a full week configuring Alibaba's Qwen3-27B INT4 model to run on two RTX 3090 GPUs connected via PCIe without NVLink under Windows 10 and WSL2. Initial attempts with vLLM failed due to CUDA library mismatches and zero speedup from speculative decoding across the two GPUs. Switching to SGLang required matching CUDA 13.0 precisely, disabling custom all-reduce, and patching a SymmMem crash specific to the 3090 architecture on WSL2. Most speculative decoding methods hurt performance, but using DSpark — an independent 1.4B draft model — boosted throughput from a baseline of 63 tok/s to 170–210 tok/s by achieving roughly four accepted tokens per verification pass.