Ran TensorFold (EXL3) against my own vLLM TP2 recipe, same prompts:
⥠Code 82 vs 69 tok/s
⥠JSON 84 vs 71
⥠Prose 44 vs 35
đ Mixed load at C4: 108 vs 72 tok/s (+51%)
đ Prefill 1,781 vs 1,225 tok/s
đ§ 500K context vs 262K
Two boxes on a desk đ¤¯