Every public OKC solution I found, what to take from each, and the corrections the deepest data forced on this plan. Updated Nov 14 after a second research pass.
| Repo | Standing (their report) | Take from it |
| keell0renz | decode #38, bf16 #26, prefill #29, attention #35, MoE #58 (Oct 31). Best public on all five boards. | Start here. Code for all five problems. 12-bit BF16 codec, 6-bit scales, stream-K wgmma MoE, per-M builds, PDL trigger placement. LESSONS.md has about 460 experiments, failures included. |
| alygg77 | overall #40; MoE 2.40, decode 2.21 | The best measurement corpus published: docs/H200_MEASURED_FACTS.md and docs/NEGATIVE_RESULTS.md (summarized below). KPERM3 scale-sharing re-layout. |
| vjeux | the base most teams forked | Native CUDA, one file per problem. Origin of the lossless BF16 weight format. Its prefill and attention files were still near the top unchanged. |
| yesthatwilly | #147 overall; prefill #50 | Closing-board leader numbers and rules. A 792-line claims ledger (landed / lead / shelf / refuted) per problem. Their verdict: the public keell0renz and alygg77/vjeux files are the fastest known, and own work added 0.1–2% per row. |
| georgxzh | mid-board | Per-shape SGLang µs for every problem, the rule-change timeline, and the DeepGEMM path for 16K prefill experts. |
| Kushagrabainsla | mid-board | Clear designs: cluster/DSMEM attention, PDL MoE, norm folding. Lessons on cold code, compiler-sunk loads and PDL __ldcg. |
| hariappi | attention 2.657×, MoE 2.381×, bf16 1.584× | A hybrid attention dispatcher (own kernel for full layers, keell0renz for sliding). Proof that mixing per-layer-type kernels pays. |
| trxphxle | #191 overall | Restoring SGLang's BF16 rounding points. Exact KV gather-read +0.63%. Pre-warming +0.16%. |
| lee101 | lower-mid | Diagnosed full verify as CTA-barrier bound. Ones-wgmma row sums. Heaviest-first MoE queue. Router [hi;lo] stacking. |
| ruwwww · tech-cosmos · shashank-100 · infraax · maximenewman · chenwainuo | various | tech-cosmos: Gluon warp-specialized full-attention kernels (2.54×, rank 8 early on). infraax: an AI-agent campaign with a "predict, then kill" method. maximenewman: clear H100 per-shape tables. chenwainuo: serving-track requirements and patches. |
Not public: ZhimaoL's sglang fork branches (decode #29), which were deleted Nov 1; their verify kernels survive inside keell0renz's files. Closing #1s (pjmullin3 attention 3.339×, gogowhy decode 2.577× at one point, dolanzhao MoE 2.758× on Oct 31, arashattar bf16, zzz0906 prefill) have published no code that I could find.
| Source | What it gives | Caveat |
| Official rules summary (llms.txt) | The serving contract: v0.5.23 base, lossless, score definition, loads, TTFT and memory gates, patch scope and the 400 KB cap, maintainer review, 1 scored run and 3 checks a day, the prize split. | A summary. The full rules page and the serving API need a browser or the CLI. |
| chenwainuo/okc | The only public serving repo found. Active patches: small-batch fused finalization, and Hopper EP8 expert-tile caps (16 through 256 tokens, 32 through 512, stock selection above). Validation gates: baseline, profile, hypothesis, isolated change, paired measurement, re-profile, full 8-GPU check. Earlier routing candidates were parked for showing no useful gain. | No serving results published. The speedup ranges in its rewrite plan are labelled priors. |
| SGLang low-latency mode (Oct 18) | What the maintainers themselves ship for low concurrency: dependent kernels launched early with PDL, no separate read-back of the attention projection output, less routing work in MoE, draft sampling combined into one kernel per GPU. Per-user decode 12–38% faster at 1–8 requests. | GB300 and a different model, delivered as an external plugin. A guide to what they consider reasonable, not a number to expect on H200. |
| LMSYS megakernel post | About 1,450 launches per stock step, 12.9 → 5.9 ms, draft 1.16 → 0.44 ms, sync and fence costs. | Batch 1 only, Blackwell only, 8,000 lines. |
| SGLang cookbook | Hopper specifics: Marlin W4A16 experts, static memory fraction 0.85, no MegaMoE or sequence-parallel prefill. DFlash2 with 8 draft tokens in the Blackwell recipes. | No complete H200 launch line. Its high-throughput profile drops speculative decoding, which this track does not allow. |
| Quesma: reports from the top 50 | The closing spread (#1 / #10 / #50), how twelve top-50 teams worked, costs, and where profiling counters are available. | Self-reported; kernel round only. |
| Fact | Value | Implication |
| DRAM read ceiling | 4.60 TB/s on a 2 GiB stream, but size-dependent: 4.33 at 256 MiB, 3.63–3.81 at 66 MiB (NCU) | Per-call streams of 50–600 MB can't reach 4.6. Use size-matched ceilings for floors. |
| Per-SM HBM pull | bare TMA ring 140–207 GB/s; plain 16 B loads 111–125; consume-before-refill ring ~60–64 | Never let consumers gate refills. |
| Per-SM write to L2 | ~40–50 GB/s (a 128 KB FP32 tile takes 2.6–4.1 µs) | Why split-K partials lose at m256 and in big merges. |
| Integer ALU vs HGMMA | ALU alone 286 clk, MMA alone 256, together 489 | Dequant does not overlap wgmma on the same SM. Instruction count per weight is the MoE and codec lever. |
| Empty graph replay | 4.70–4.74 µs on their timer | The judge's ~4 µs subtraction leaves a bit of fixed cost in each shape. |
| Decode layer budget, d32s (154 µs) | MoE 84 · dense GEMV 51.5 (floor 34.9) · glue 8.8 · attention 9.8 | MoE is over half the layer; dense has ~16 µs of headroom. |
| Decode layer budget, v32s (306 µs) | MoE 188 · dense 83 (floor 72) · attention 14.8 · glue 18.4; idle between kernels <1.1 µs | Inter-kernel bubbles are already small on the verify path. |
| MoE loss budget (WGS build) | ramp 3–6 · phase-A waves 3–6 · shadows 4–12 · end 7–11 µs. Stream phase 3.87–4.05 TB/s. | Matches keell0renz: the losses are the edges, not the stream. |
| Phase-B unit vs end spread | a unit is 18.5–19.4 µs; end spread is 2.7–8.2 µs | Dynamic claiming can't fix the tail. Make the last units smaller. |
| NVFP4 entropy | E2M1 codes 4.000 bits; scales 3.71 bits (34 codes) | Codes are incompressible; best lossless total saving is ~5%. |
| BF16 weight entropy | 10.46 bits; escapes 1.3% at 12 bits | Entropy coding below 12 bits: upper bound +2.4% even with free decode, realistically +0.5–1%. |
| Full verify consumer loop (v16f) | exp 40.5% · QK + V transpose 17.5% · PV barrier 12.7% · data waits 12.5%. 1 CTA per SM is 16–18% slower than 2. | exp offload is the lever for full verify. Keep 2 CTAs per SM and a 104-register budget. |
| Verify attention beside a GEMM | v32s 14.9 µs alone vs 22–27 µs co-running | Don't overlap attention with weight streams. |
| Prefill first-turn attention | 12.3% of the first-turn load; a kernel with no exp still takes 87–91% of the time | FP8/exp tricks have a small ceiling for the first-turn load. |
| Prefill glue | 1,108 µs memory-bound glue (13.5% of load); GEMMs at 1.53 GHz, slowed 12–18% by other traffic | Fusing glue into GEMM epilogues beats hiding it under GEMMs, which lost. |
| prepare() time | 273–275 s typical, once 311.9 s (limit 300) | Variant sweeps can time out. Budget compile time. |
| Judge noise | the same file spread 1.6–2.2% across runs; cross-cluster comparisons inflated a +2.4% gain to +4.19%; memory-layout pads shift v32 by 1.5–2.5 µs | Compare only within one cluster, ≥2 runs per arm, and test ≥4 layout pads. |
| Page | Earlier claim | Correction |
| overview, serving | Serving runs Nov 10–19; rules taken from public notes | Nov 12, 9 am PT to Nov 19, 11:59 pm PT. Since Nov 11: SGLang v0.5.23 base and lossless patches. 1 scored run and 3 checks a day. |
| overview, serving | Kernel work transfers almost one-to-one; every kernel can drop into SGLang | Withdrawn. 400 KB of python/sglang source, +2 GB per GPU of build memory, nothing below stock precision, and a maintainer review that can disqualify load-specific code. |
| overview | Phases 0–6: 10–13 days of kernel-board work, then serving | Reversed. The boards are closed and unpaid. Serving comes first; the kernel pages are a parts catalogue. |
| serving | DFlash tuning per concurrency, fused all-reduce and N-sharded experts ranked high or medium | Downgraded to check-first or out: fixed flags, load-level special-casing, patch size and review cost. |
| attention_decode, prefill_layer, bf16_proj | FP16 accumulation, emulated exp2, FP8-stored P, hi/lo BF16 router | Kernel-track allowances only. Assume out in serving until RULES.md says otherwise. |
| bf16_proj, moe_w4a16 | Re-encode weights and keep per-shape builds | Qualified. One layout per weight, converted in place. About 32 MB of duplicated weights per layer is all the memory cap allows. |
| Page | Earlier claim | Correction |
| decode_layer | Router-first / pooler-concurrent reorder and o-proj shared-memory prefetch during attention | Downgraded. An early pooler/router stream at d32s measured 2.1–7.7 µs slower; L2 prefetch from sliding attention was 2.6–5.1 µs slower; 2 pre-wait weight stages are optimal and 4–6 slow attention by 1.7 µs. Only an in-launch variant is still untested. |
| moe_w4a16 | Dequant into wgmma A registers "fixes" instruction-bound large-M shapes | Qualified. Integer ALU and HGMMA don't overlap, so the win must come from fewer instructions per weight (KPERM: 20.12 → 18.44 per word; −12 instructions per k-block gave +5.8% on prefill experts), not from overlap. |
| bf16_proj | Entropy-coded ~11-bit exponents; stream-K for qkv | Downgraded. ≤+2.4% upper bound (realistically +0.5–1%). qkv GEMV saturates at ~96 CTAs; stream-K's ceiling is 0.6 µs vs a 1.5–4 µs fix-up. 12-bit decode at m≥128 loses 2.4–12 µs. |
| prefill_layer | FP8 attention as a top lever | Reduced to medium. Attention is only 12.3% of the first-turn load. Top levers become the expert kernel at T=753 (423 TFLOP/s, ~25% padding, instruction-bound) and fusing the 1.1 ms of glue. |
| attention_decode | Emulated exp2 only at v32 full | Strengthened. exp is 40.5% of the full-verify consumer loop at v16 too. Try it on v16f and v32f. |
Couldn't read (GitHub rate limit): yesthatwilly's per-problem CLAIMS.md ledgers. These are the next sources to mine. Also unread: each repo's per-experiment files (Kushagrabainsla EXPERIMENTS.md, infraax HYPOTHESES.md, trxphxle full tables).