{"model":"deepseek-4.1-flash","title":"KV accounting measured (PP8)","status":"positive","summary":"Zanooda PP8, fp8_ds_mla: 6,170,576-token KV pool in 135.72 GiB total (per-rank 6.48-20.95 GiB) -\u003e 23,617 B/token whole-model; 1M tokens ~23.06 GiB (~26x the model card's 890 B/tok) because fp8_ds_mla keeps uncompressed indexer-K caches. Goal reprioritized to PP8/PP6 only.","occurred_at":"2026-09-12T04:22:00Z","config":{"serving":"TP1xPP8 (Zanooda 13-patch stack), fp8_ds_mla","engine":"zanooda/vllm-sm80-ds41f:v41-sm80 + KV-layout intersection fix"},"benchmarks":[{"metric":"kv_pool_tokens","value":6170576,"unit":"tokens","context":"PP8 pool (5.88x concurrency @1M)"},{"metric":"kv_gib_per_1m","value":23.06,"unit":"GiB","context":"PP8 (23,617 B/token) - vs 10.95 GiB on PP6"}]}