Pull requests / #2045
#2045 bench: RTX 3070 IQ3_S MTP completion-time crossover
open · @CC-David-CC · 0 コメント · GitHub で見る
BenchmarksNVIDIA / CUDAModels & quantsDocumentation
本文
Adds an exploratory RTX 3070 8 GB / Ryzen 7 5800X report for Flash-Next GSQ-RCO IQ3_S with INT8 KV, measured on pinned build `61b3fb5d`.  - Recorded sustained crossover: about 522 output tokens at 4K input, and 1,426 at 16K. - At 64K input, the off arm ended at 759 tokens after repeated imports; no long-output crossover is claimed. - One run per condition, warm engines, no prompt reuse; MTP and `--spec 2 -> 4` change together, and output IDs diverge. Includes the [full report](https://github.com/CC-David-CC/Strata-a5500/blob/7d50763b89bc67b748abd53aaa1a9be8e887b49e/bench/results/2026-10-12-community-rtx3070-mtp-crossover/README.md), raw arrivals/IDs/outputs/logs, mathematical model, reproduction harness, PNG/SVG chart and plotting script. Adds a community-report index entry. Validation: evidence arithmetic, output hashes, observed crossovers, early-stop handling, links and Python syntax passed; chart visually checked; `git diff --check` passed. Documentation/results contribution; no engine or default changes.
関連リンク
インストール・モデル・リリースへの站内リンク。