Pull requests / #2604
#2604 docs: STRIX_HALO.md: the first Windows report, and STRATA_ARENA_PIN_GIB=16 that made it start (#2013)
closed · @qiuzan2001 · 0 Kommentare · Auf GitHub
BenchmarksSetup & installAMD / HIPModels & quantsDocumentationWindows
Beschreibung
The docs part of #2013. **What it changes.** `docs/STRIX_HALO.md` said the Windows zip's gfx1151 code was untested on a Windows Strix Halo. A short "Windows (one report)" section (linked from that sentence) gives #2013's result with the reporter's numbers: Ryzen AI Max+ 395, 128 GB, 64 GB BIOS carve-out, Windows 11 Pro, AMD driver 32.0.31036.15, the 0.1.43 ready-made zip, IQ3_S at 131K. The engine never became ready: page-locking the whole ~47 GiB arena took about twice its size from the GPU's free memory (101,470 -> 5,540 MiB) and the prompt path then failed on its first GPU call (`unspecified launch failure`). `"STRATA_ARENA_PIN_GIB": "16"` in the run config's `"env"` (pins at most the first 16 GiB, in whole layer slices: 15 GiB there; the rest locked resident) fixed it: ready in 33-37 s, all 24,576 experts cached, 44.7 / 45.0 tok/s on a 256-token answer, 708-733 tok/s on 4K-32K prompts (`strata_bench.py --long`). Other values were not tried. **What it leaves alone.** The code and setup's defaults (a cap on Windows UMA by default would be a separate, measured change), and the rest of the page. **How it was checked.** Against #2013 and its comments, and `arena_pin_cap_gib()` / the `PinnedArena` constructor in `src/core/pinned.cu` for what the variable does. A sub-agent reviewer checked every number; its corrections (the failure wording, what the variable pins) are in. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Mehr auf der Site
Links zu Install, Modellen, Releases.