A 13-day kernel/scheduler diary is mapped layer-by-layer onto a 7-layer OSI model to identify where each optimization contributed to a 3.22× end-to-end throughput gain over a W8A8 baseline.
Adapted from @alexdolbunQ4 2026 · @alexdolbun & “OSS AI Agentic Swarm Orchestra Conductor / 智能体蜂群乐团指挥”:> @Zai_org 的 GLM-5.3，出自唐杰教授 @jietang —— 清华计算机系教授、https://t.co/9CEidd6Wvs 创始人（HKEX: https://t.co/jeEzCJFB29）—— clearly a huge leap after Mr. @finkd & Mr. @alexandr_wang $META AIRA₃ the $NVDA by Mr. @JensenHuang backed “NVIDIA Kaggle Nemotron Model Reasoning Challenge”, where Meta's AIRA₃ multi-agent parallel processing or dApp dDojo environments system secured an 8th-place gold medal finish out of roughly 4,000 competing human teams. The system evaluated in that swarm setup combined frontier models including GPT 5.5, Claude 4.8, and Meta's own Muse Spark… https://t.co/xBDj2QGolY … The screenshots confirm the Infra Agent loop. Next I’ll pull the rest of the post plus the actual kernel PRs, weights, and the DeepEP/Mooncake path so the stack can be mapped layer-by-layer.The paper is not an RSI manifesto. It is a 13-day kernel/scheduler diary on a hybrid KDA+DSA MoE, with the control plane written by GLM-5.3 and the data plane stuck on immature Chinese-accelerator kernels. The alpha is in three places: (1) the day-by-day multiplier curve, (2) the Python/C++/RDMA seam, (3) the OSS that already absorbed the patches. OSI map of this serving stack Treat the cluster as a 7-layer machine. Every bug they published sits on a specific layer. That is the extractable part. OSI In this system What actually moved 3.22× L1 Physical >100k Chinese accelerators, tight HBM + HBM bandwidth, multimodal + 1M ctx Everything below is a reaction to “not enough bytes/s off chip” L2 Link GPU-Direct RDMA, PCIe topology → NIC affinity (Mooncake auto-detects) SM-free KV move. Compute SMs stay on GEMM/KDA L3 Network DeepEP expert-parallel dispatch/combine (intra vs inter node); Mooncake TE batched read/write Expert tokens and KV/state are two different fabrics on the same NIC L4 Transport Mooncake Transfer Engine + (in SGLang recipes) NIXL; Python GIL as a process-local mutex on the control plane This is where they found the 20% → <1% bug L5 Session EPD = Encode / Prefill / Decode disagg. Session state = paged DSA KV + KDA recurrent state PD is not “just KV”. Hybrid models ship a second tensor L6 Presentation W8A8 weights, INT8/FP8/BF16 KV, tf32 vs tf32x3 on Triton https://t.co/E6jZ8cW2mF, Layer Split, IndexPool Numerical format is the protocol L7 Application SGLang router + Infra Agent + dense-feedback harness Agent only works if L6/L4 emit local oracles If you copy one sentence into a swarm conductor: EPD handoff for Flash is (DSA paged KV, KDA SSM state). Drop either and decode is silently wrong. SGLang’s own cookbook already says PD “moves both the paged DSA KV and the KDA recurrent state.” Figure 1 is the real paper From the chart you screenshotted, end-to-end gain vs W8A8 baseline: Day Change × T+0 W8A8 baseline 1.00 T+1 Async scheduling 1.21 T+2 Sort kernel 1.42 T+3 Hierarchical cache 1.41 T+4 Layer Split 1.97 T+5 Context parallel 2.49 T+6–8 KV-transfer overlap + mixed-precision cache quant + chunked MQA 2.67 T+9 Prefill dequant kernel 2.85 T+10 Fused activation + quant 3.01 T+11–13 launch Linear attention path 3.22
