Pull requests / #3658
#3658 [Benchmark] RTX 5080 / Qwen3.8-Flash-Next / UD-IQ4_XS / strata-qwen38-thinking-v1-draft.1
open · @Teracon00 · 0 コメント · GitHub で見る
BenchmarksSetup & installAMD / HIPNVIDIA / CUDAModels & quantsDocumentationWindows
本文
Standard benchmark report for RTX 5080 / Qwen3.8-Flash-Next / UD-IQ4_XS (unsloth), protocol `strata-qwen38-thinking-v1-draft.1` (draft).
**Run:** `20261021T214719Z-af066394` — folder `bench/results/2026-10-21-standard-rtx-5080-af066394/` (`README.md`, `results.json`, `outputs.jsonl`, `config.json`, `engine-log-excerpt.txt`, `provenance/`).
**Status:** complete. 21 of 21 required requests finished, 0 failed, 0 skipped; sanity 12/12 (instruction 4/4, retrieval 4/4, tool calls 4/4). Optional concurrency stage not run. Time after model load: 2 min 39 s.
**Setup:** Strata v0.2.2 (commit `e6fff4ad12908de74952abe53b2bade45a68aa0f`), engine 0.1.47 release build, unsloth UD-IQ4_XS pack, context 131072, KV q4_0, conversation cache 8192 MiB, VRAM reserve 1024 MiB. Run against a live server via `strata bench run --suite standard-v1 --url http://127.0.0.1:8080`.
**Hardware:** NVIDIA GeForce RTX 5080 (15.9 GiB), AMD Ryzen 7 7800X3D, 94.9 GiB RAM, NVMe (INTEL SSDPED1D480GA), Windows 10.0.26300, gpu-driver 617.42, PCIe Gen4 x16.
**Limitations:** expert arena pinned with 4 KB pages (large pages refused by the OS in all loads); expert cache auto-sized to 3074 slots (6.92 GiB), decode hit rate across the run ~67 %. Model/source/tokenizer revisions, runtime and background work are `null` with reason in `results.json` (live server does not expose them).
**Note:** an earlier attempt the same day (`a558f5d5`, 21:15) was nonconforming ("the Strata commit is unknown" — no COMMIT file) and is deliberately not submitted. `af066394` is the conforming UD-IQ4_XS run for this configuration.
関連リンク
インストール・モデル・リリースへのサイト内リンク。