This folder is the promoted and closed-out result ledger for the lab. It is not limited to the currently active model. Use it for model-specific outcome packets, record summaries, validity gates, reproduction commands, and links to evidence.
Keep active experiments in ../experiments/, chronological investigation notes
in ../notes/, patch snapshots in ../patches/, compact run evidence in
../data/, and copy-ready recipes in ../repro/.
For a compact, manually maintained view of expected performance across model and hardware identities, see the performance index.
Historical first-100-token rows from the shared realistic-suite helper used a 100-event numerator over a 99-interval span. They remain preserved as published evidence; use the scoreboard accounting note and model-specific corrections before comparing their absolute values with conventionally counted rates.
| Model / Lane | Folder | Status | Best Valid Result |
|---|---|---|---|
| Poolside Laguna S 2.1 INT4 on 4x B70 | qualified packet; repro; correction; structured record | Approved and sealed; metric qualified | Four-B70 TP4+EP4, one active generation, exact target-verified DFlash depth 11 on the audited width-12 Breakable PIECEWISE graph with 31 FP8 W8A16 draft projections per rank: 102.971436 tok/s under the submitted legacy 100-event/99-interval convention and 101.941721 tok/s under conventional interval accounting. First valid cold score; 13/13 token-and-text exact vs canonical q1, all cached_tokens=0, 512-output-then-next 2/2, rollover 1/1, no warmup or retry; LocalMaxxing cms2ccv2d00lps201rej94pjy. |
| DeepSeek V4 Flash uniform-K160 on B70 | result packet; standalone repro; closeout | Paused/closed frontier | Four-B70 TP4+EP, one active generation, unchanged K160 target with target-verified DSpark7: 80.820052 tok/s strict high and 78.287226 tok/s three-suite median-of-medians; 36/36 cache-zero realistic rows, 24/24 exact canaries; LocalMaxxing cmrquta9905w3lg013m5vxoqx. No later verified endpoint exceeded it. |
| Qwen3.6 27B INT4 AutoRound on B70 | handoff; packet; TP2 record; repro | Closed reference lane | Two-B70 TP2 vLLM/XPU strict result: 95.384868 tok/s median for generated tokens 1-100 after TTFT on the fixed realistic cold suite, target-verified MTP3, cached_tokens=0, with complete exact/repeat128/baseline/needle quality evidence; LocalMaxxing cmrh35ct50092mj01h7jgydqj. |
| Qwen3.6 27B GGUF Q4_0 native DFlash/SYCL on B70 | closure; experiment packet | Closed intensive research lane; objectives missed | One-B70 llama.cpp/SYCL, unchanged Q4_0 target with target-verified native DFlash5 and fused Xe2 verifier/draft-head stack: 47.818818 tok/s median for generated tokens 1-100 after TTFT on 12 unique cold prompts, cached_tokens=0; LocalMaxxing cmrjbx8bc02g8mj01yzz2v701. The requested 100 tok/s TP1 and 200 tok/s multi-B70 single-session objectives were not reached. |
| Qwen3.6 27B MTP GGUF Q4 on B70 | packet | Valid reference, not competitive | One-B70 llama.cpp/SYCL, unsloth/Qwen3.6-27B-MTP-GGUF, Qwen3.6-27B-UD-Q4_K_XL.gguf, draft-MTP3, fixed realistic cold suite, cached_tokens=0: 30.678766952807752 median tok/s for generated tokens 1-100 after TTFT; LocalMaxxing cmr6mn5ct0076mn01on3dnpyn. Useful as a model/runtime variation reference; the AutoRound vLLM lane is the faster Qwen27 INT4/Q4 headline. |
| Gemma 4 26B A4B Q8 / INT8 on B70 | handoff; production service; packet; strict repro: ../repro/gemma4-26b-a4b-q8-b70-125tps-20260701 | Production-servable frontier/reference | One-B70 llama.cpp c926ad098, UD-Q8_K_XL target/verifier with Q4_0 MTP draft verified by target, fixed realistic cold suite, cached_tokens=0: strict high 124.97714084813418 median tok/s for generated tokens 1-100 after TTFT; LocalMaxxing cmr1u77na01k2ld01kalwzs1e. Same-family support and variance notes are in the packet; synthetic/warmed rows are diagnostic only. |
| Qwen3.6 35B A3B Quark W8A8 INT8 on B70 | qwen36-35b-quark-int8-b70 | Closed reference packet | 4x strict-valid 93.55 tok/s; TP2 safe smoke 85.87 tok/s |
| MiniMax M2.7 INT4 AutoRound on B70 | result packet; deploy repro; strict repro | Deployable baseline plus historical strict and constrained-task lanes | Deployable 32K endpoint: 83.172 output tok/s on the comparable p512/n1536 gate. Historical strict 2K lane: 89.314195 output tok/s. Constrained structured-output results are labeled separately in the packet. |
| Gemma 4 12B IT INT4 AutoRound | ../experiments/gemma4-12b-int4-autoround-vllm | Production slot plus research profiles | c8 32K image+text endpoint; c10/c12/c16/c64 documented separately |
Every row above was measured in the reference B70 lab. Contributed results that have not been reproduced here are listed separately below and are not part of this ledger.
Outside contributions that have not been reproduced in the reference lab.
These are not lab results and are not comparable to the table above: the
numbers are as reported by their contributor, on their hardware, under their
methodology. Entries live in ../community/ and each
one carries a STATUS.md recording exactly what was and was not checked here.
A row graduates out of this table into the ledger above only after it reaches
B70-tested or better, per
../docs/contribution-verification.md.
| Model / Lane | Entry | Evidence level | Contributor-reported claim |
|---|---|---|---|
| Qwen3.6 27B FP8 native TP2 Docker on 2x B70 | community entry; PR #9 | B70-tested; recipe executed here 2026-07-25 |
Recipe confirmed working in the reference lab: TP2 serves Qwen/Qwen3.6-27B @ 6a9e13bd6 at native FP8 under intel/llm-scaler-vllm:0.21.0-b1, 14.13 GiB/card, KV cache 888,488 tokens, healthy ~110 s after start. Measured 30.171 tok/s median decode (range 29.564-30.528, stdev 0.302, 15 rows over 3 passes, one active generation, 256-token outputs), which falls inside the contributor’s reported 28.3-34.7 range; cite this figure rather than their top-of-range “34”. Card pairing across PCIe root complexes made no measurable difference (30.482 median). Sampling enabled per the contributed config, so no exactness gate; cached_tokens not reported by this build, prefix caching disabled and prompts unique. The multi-GPU Level Zero failure reported by a second contributor does not reproduce here. |
Promote a result into this folder only when:
../data/ or an external path is
recorded;If a model lane is paused or closed, keep the packet. Paused packets are how future work avoids rediscovering the same dead ends when we switch back to a model after learning something from another lane.
For LocalMaxxing submissions, see ../docs/localmaxxing.md.