Full Kimi K3 running on 16x GB10 cluster

The new V3 image with further long context optimization and fixes is now live:

From llama-benchy (coherent corpus, tg=1500, up to 300K context). General-purpose corpus, so DSpark acceptance is on the lower side:

test t/s peak t/s ttfr (ms) est_ppt (ms) e2e_ttft (ms)
pp2048 @ d4000 675.10 8,260.59 8,256.52 8,260.59
tg1500 @ d4000 19.82 35.00
pp2048 @ d50000 748.43 62,921.28 62,917.21 62,921.28
tg1500 @ d50000 20.09 39.00
pp2048 @ d100000 712.79 129,595.93 129,591.86 129,600.92
tg1500 @ d100000 17.41 37.00
pp2048 @ d200000 657.64 277,772.00 277,767.93 277,780.66
tg1500 @ d200000 15.66 33.00
pp2048 @ d300000 607.73 449,404.11 449,400.05 449,417.24
tg1500 @ d300000 14.85 28.00

Single-stream coding benchmark (temp=0, prompt 125 tok, tg=1500) — coding workloads see higher DSpark acceptance than the coherent corpus above:

Completion tokens : 1500
Prompt tokens : 125
Wall time : 59.09s
Decode tok/s : 25.38

What’s new in v3:

  • Faster decode: ~+5-8% vs V2 .
  • Long-context loop fixed: the NaN/generation-loop bug observed past ~270K context under DCP is fixed by keeping the recurrent KDA token-position cache unsharded under DCP — a fix found and upstreamed by the local-inference-lab/vllm project as PR #418. Credit for identifying and fixing this issue goes to that repository’s authors.
  • 2M total context capacity: the TP16+DCP8 recipe holds ~2M tokens of KV — 4 concurrent requests can each use a 500K context.