Pull requests / #2621

#2621 bench/results: RTX 5070 Ti 16 GB on Linux, Flash-Next IQ2_XS as setup installs it (0.1.44.1)

open · @qiuzan2001 · 0 コメント · GitHub で見る

BenchmarksSetup & installNVIDIA / CUDAModels & quantsDocumentationWindowsLinux

本文

Results only: one new folder, `bench/results/2026-10-15-community-rtx-5070-ti-iq2_xs/`, its line in `docs/COMMUNITY_BENCHMARKS.md` and its row in `bench/results/COMMUNITY.md`. No engine, tool or setup change.

**Machine:** RTX 5070 Ti 16 GB (gen 5 card in a gen 4 x16 slot, 300 W), Ryzen 7 5800X3D (AVX2, no AVX-512), 80 GiB DDR4-3200, Ubuntu 26.04, driver 595.99.02, CUDA 13.1. Strata 0.1.44.1 (`296b59d2`), engine compiled by setup for sm_120. The original Flash-Next **IQ2_XS** at setup's pinned revision, installed by `setup.sh --setup --yes --family qwen --model IQ2_XS`: setup chose a 65,536-token context, int8 KV, 6,417 experts in VRAM.

**Results** (engine timings, medians and ranges of 3):

| | |
| --- | --- |
| `strata_bench.py --long`: chat decode | 116.0 tok/s (95.4-125.2, n=9) |
| prompt at 4K / 16K / 32K | 2,817 / 3,380 / 3,393 tok/s |
| agent session decode, cache hit | 126.7 tok/s, 68.7% |
| decode, 256 tokens after 4K / 32K / 58K prompts | 167.2 / 120.1 / 109.3 tok/s (wide ranges, with draft acceptance) |
| needles (`needle_bench.py`) | 6 of 6 at 32K and 64K |

GPU clocks, power and temperature (2,872 MHz, 260 W median) are in the folder. The same suite on another RTX 5070 Ti (#2173, Threadripper 7960X, Windows) measured 123.6 tok/s chat decode and 2,795 / 3,287 / 3,320 tok/s prompt with this size; the README names that, and why it is not like-for-like.

**What it leaves out:** one engine session of about five minutes; no 128K context (setup chose 65,536 here); no vision, concurrency or quality test.

**How it was checked:** a second reader recomputed every number from the attached data (results.json, decode-sweep.json, needles, engine log, telemetry, free -m and the hashes) and searched all files for private data. The model shards' SHA-256 match the repository's. This is the third report from this PC (with #2580 and #2617); all three add a line at the end of the same list and a row at the end of `bench/results/COMMUNITY.md`, so the later ones have a one-line conflict there.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

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。