For AI agents: the complete documentation index is available at https://a3s-lab.github.io/Power/v1.0.0/en/llms.txt, the full documentation bundle is available at https://a3s-lab.github.io/Power/v1.0.0/en/llms-full.txt, and this page is available as Markdown at https://a3s-lab.github.io/Power/v1.0.0/en/performance.md.
A3S Power
  • English
  • v1.0.0

      • Performance evidence

      Power accepts performance work only when the measured system still preserves its execution contract. The published Qwen3.8 captures therefore use the real streaming API, pin artifact and binary identities, verify deterministic output, and report quality and workload throughput beside steady decode.

      This page documents one backend/model/hardware integration, not the scope of the engine. Power's shared execution, admission, memory, cancellation, and evidence contracts are model-neutral and also serve vision, OCR, embedding, audio, multimodal, scientific, and caller-owned reviewed graphs.

      To verify the numbers directly, open Reproduction: run the model-free one-command verifier first, then replay the full RTX 4090 protocol against the same input identities.

      Quality and speed by mode

      Active Qwen3.8-27B Q6_K modeFixed 100-task score, mean of 3 runsRequest-wide throughputThree-run range
      Untouched Q6_K, autoregressive control67/100 lenient; 60/100 strict23.642 token/s23.543--23.832
      Same Q6_K, full-vocabulary MTP67/100 lenient; 58/100 strict41.035 token/s (1.736x)40.197--41.696

      All six runs completed 100/100 requests without errors and passed continuous per-process GPU exclusivity checks. Each mode was internally deterministic. Across modes, MTP retained 89/100 extracted answers and 50/100 complete output hashes. Strict formatting moved from 60/100 to 58/100, so this faster mode is not a lossless production default.

      Request-wide includes prompt processing, generation, HTTP, and request overhead. It is not the same metric as warmed-up 1,024-token steady decode. The Q6_K-only evidence package can be checked without a model or GPU.

      Archived auxiliary-proposer research

      The DSpark rows use a shorter context-512, 256-token acceptance shape rather than the 1,024-token MTP peak shape. They are paired with each other, not with the historical autoregressive row.

      Q6_K-only native DFlash2 boundary

      Both modes keep the 22,884,408,288-byte Q6_K target unchanged. The 1.14 GB Q4 DFlash2 artifact only proposes tokens; Q6_K verifies every committed token.

      Same Q6_K targetMedian decodeMinimum decodeMedian end-to-end
      Target-only33.075 token/s32.938 token/s25.744 token/s
      DFlash2 K7/S6144.453 token/s141.267 token/s63.182 token/s

      That is a 4.367x decode and 2.454x end-to-end speedup. All five paired outputs matched, acceptance was 98.230%, and replay was zero. The representative 12-task quality calibration retained 12/12 extracted answers but only 7/12 complete outputs. DFlash2 is therefore a native opt-in profile, not a lossless default or a stable 175 token/s floor.

      Native external-DSpark boundary

      Power's verified external-draft path keeps the target Q6_K bytes unchanged and loads a separate 1.10 GB DSpark Q4 proposal model. Fixed K10/S6 produced 169.561, 167.102, and 169.324 token/s, a 169.324 median and 5.250x gain over its 32.249 token/s paired target-only control. Proposal acceptance was 90.873%, verified tokens per target pass were 9.8077, and replay was zero.

      Every paired request, 256-token output, and execution receipt had the same digest. That is exact evidence for this deterministic request, not a cross-prompt intelligence result. The profile peaked at 23,847 MiB and left only 717 MiB on the recorded RTX 4090, so low utilization alone does not prove that model admission will be stable.

      The separate 600-request quality capture used 100 fixed MMLU/GSM8K/C-Eval tasks, three repetitions per mode, context 1,024, and batch 12. Target-only scored 67/100 lenient and 58/100 strict at 22.618 token/s request-wide. DSpark scored 73/100 and 59/100 at 32.678 token/s, a 1.445x gain. Both modes were deterministic, but only 54/100 complete outputs matched across modes and every DSpark request entered exact fallback replay. The observed score did not fall; nevertheless K10/S6 fails the lossless production-default gate and remains an explicit benchmark profile.

      DFlash v1 is not part of this number. It uses a different artifact contract and cannot be layered onto DSpark. No genuine DFlash v1 GGUF has completed the acceptance gate on this host; DFlash2 is the separate native capture above.

      What the pure-Q6_K peak boundary means

      The latest exact-build replay of the untouched 22,884,408,288-byte Q6_K artifact reached 174.413 token/s median across nine 1,024-token samples with prefix-FR8192 K7/S6. Its minimum was 172.723 and every output digest matched, so it did not establish a stable 175 token/s floor. An earlier quiet-host capture reached 176.6109 token/s median with a 173.263 minimum; neither is a service guarantee.

      No model weight was requantized. The speedup comes from native MTP, exact target verification, an 8,192-row draft-only token-ID prefix, six resident recurrent snapshots, batched GPU greedy sampling, normal CUDA Graphs, short-batch Flash Attention off, full CUDA offload, batch 11, and host controls.

      This is a high-coverage peak, not a service floor or universal default. On the one-pass 12-task calibration, full-vocabulary K7/S6 reached 47.032 token/s request-wide with 52.30% acceptance, while prefix FR reached 37.290 token/s with 24.82% acceptance. Eleven tasks per mode hit the output cap.

      The previous mixed-artifact boundary remains separate. Its rollback-complete K7/S7 profile reached 175.2089 token/s steady decode and 83.228 token/s on the repeated 100-task workload. That 19,187,686,464-byte artifact uses Q4_0 main FFN tensors, a Q6_K MTP block, and a Q4_K draft head.

      The current acceptance host is Windows 11 with an RTX 4090 and a 10-core, 20-thread Intel Xeon w5-2445. Its 0x55555 affinity mask is topology-specific, not a portable product default.

      Did quality fall?

      The active Q6_K-only matrix kept the lenient score at 67/100 while strict format scoring moved from 60/100 to 58/100. MTP had two lenient gains and two losses, 89/100 extracted-answer parity, and 50/100 complete-output parity. All 59 tasks untruncated in both modes retained the same extracted answer. This does not prove unchanged intelligence or lossless output, so the default gate remains closed.

      The external-DSpark matrix observed no score decrease: 73/100 lenient and 59/100 strict versus its target-only control at 67/100 and 58/100. All 58 tasks that were untruncated in both modes retained the same extracted answer. The 54/100 complete-output parity still blocks a lossless-default claim; fixed-task scores are not a measure of general intelligence.

      The DFlash2 calibration retained the same 9/12 score and all 12 extracted answers in both modes, but only 7/12 complete outputs matched. That small sample does not show a score regression; it also does not prove general intelligence equivalence or lossless output.

      For the previous mixed-artifact K7/S7 profile, no regression was observed on the fixed repeated sample:

      • TBQ4 autoregressive scored 70/100 lenient and 64/100 strict.
      • Full-vocabulary K7/S7 scored 76/100 lenient and 66/100 strict.
      • All 900 requests completed without errors.
      • Every prediction was stable across three repetitions.
      • Proposal acceptance was 51.33%; replay and guard activation were both zero.

      These are fixed-task accuracy proxies, not general intelligence or IQ scores. The paired differences did not reach conventional statistical significance, so the result does not prove that MTP improves general model intelligence.

      The archived prefix-FR mode is useful negative evidence. Its steady peak was high, but mixed-workload acceptance fell to 25.55%, request-wide throughput fell to 27.951 token/s, and C-Eval acceptance was only 14.21%. Full-vocabulary K7/S7 removes that draft-coverage bottleneck.

      Reproduce the boundary

      The dedicated Reproduction page provides copyable commands, the fixed environment, input SHA-256 values, complete runner arguments, output files, and pass criteria.

      The full replay has two levels:

      1. Offline verification recomputes archived report statistics, hashes, thresholds, and output identities without loading the model.
      2. Performance replay rebuilds the pinned CUDA profile, verifies the exact artifact, prompt, ACL, and binaries, then runs Power's streaming API workload.

      The current peak gate fixes one warm-up, nine measured requests, 1,024 generated tokens per request, batch 11, greedy sampling, and a deterministic output SHA-256. The runner fails on a changed model identity, wrong backend, short output, non-deterministic digest, missing host control, or missed median gate.

      Use the repository's checked-in guide and raw evidence:

      Treat a replay on different silicon, driver, display load, clock policy, model bytes, or prompt as a new result rather than silently combining it with this acceptance capture.