改动 / #3054

#3054 Community benchmark: RX 7800 XT (gfx1101), Windows 11, Flash-Next IQ3_XXS (0.1.41 release + private file-cache A/B)

closed · @solomon159753-oss · 0 评论 · 去 GitHub 看

BenchmarksAMD / HIPModels & quantsDocumentationWindowsLinux

说明

Measured on 2026-10-16 and 2026-10-18 by solomon159753-oss on one native-Windows desktop. As far as I can see this is the first **Windows** report from a **gfx1101 / RX 7800 XT** card in `bench/results` (an earlier Linux validation of this card is #254 in `docs/AMD_HIP.md`). It has two parts, and they must not be read as one before/after pair:

- **A. Strata 0.1.41 as released** (prebuilt Windows HIP engine), Flash-Next IQ3_XXS, 64K context, fresh prompts of
  about 1K / 8K / 24K / 60K tokens, three runs each.
- **B vs C. A private build, same binary, one opt-in switch off/on.** A same-day A/B of a Windows file-read change
  (`STRATA_R6_FILE_READ_CACHE`, proposed separately in an issue) at a 60K fresh prompt, three runs each.
  **This build is not a Strata release**: it is 0.1.44.1 with local changes, listed below.

Main limitations: one machine, one model, mostly one long-prompt length for the A/B; desktop apps were open during the
runs; GPU clocks, driver and PCIe link were not captured per run; part A and part B/C differ in date, engine, KV type and
RAM budget, so the A→C difference is a history, not the effect of one change.

Results-only submission, no engine changes.

---
Superseded by https://github.com/Niko1221/Strata/pull/3057 after moving this submission to my durikim1-cpu account.

本站相关内容

相关页面的快捷入口。