🧠 GLM-5.3 743B (Int4-Int8Mix) on 4× DGX Spark just got TWO new lanes 🚀
🟦 Lane 2: DCP4 + NVFP4 KV + DFlash2
📦 689,772-token KV pool @ 262K ctx: 2.35× our old best
📏 512K window boots: 594K pool, needle 3/3 at 490K tokens ✅
🟧 Lane 1: DCP4 + fp8 KV + DFlash2
📦 462,308-token pool @ 262K
⚡ Same decode on both: 54 tok/s counting · 40 code · 36 json · 17 prose (single stream)
📈 peak 63 tok/s aggregate (C5, mixed real prompts)
🎯 254K needle 3/3 on both lanes
🧪 NVFP4 + DCP merge checked token-by-token vs DCP1: inside run-to-run noise
⚠️ Trade-off: prefill ~400 tok/s
🙏 DCP4 + DFlash2 is
@ajclark's work, now in the repo with full credit. Thanks
@omanuke,
@wuwenthink,
@knapcio,
@brah_ddah & incoai
🔗