Pull requests / #1361

#1361 tools/hip: gfx1151 hipBLASLt 100500 tuning table (Windows ready-made …

open · @sclsgj · 0 comments · View on GitHub

BenchmarksAMD / HIPDocumentationWindows

Description

…engine)

The 0.1.40 Windows HIP engine bundles ROCm 10.2.0a20260930 (hipBLASLt 1.5.0, "100500"), so the existing gfx1151-100401 table is refused at start-up and the prompt's dense GEMMs fall back to plain hipBLAS. Add a 100500 table calibrated on a Windows Strix Halo 8060S (gfx1151) with tools/hip/tune_hipblaslt, same 90 rows and prompt shapes as the 100401 table.

Measured on that box (UD-IQ4_XS, --resident-budget-gib 8): 739-token prompt 2,337 vs 1,861 tok/s (+26%, plain hipBLAS median of 3); 2,803-token prompt 6,665 vs 6,922 tok/s (~0). Decode unchanged (CPU-pool/SSD-bound there).

docs/STRIX_HALO.md: note the Windows table and the first Windows Strix Halo
measurement. bench: tools/bench_tok.py and results.

Related on strata.com

Editorial links to help you install, pick models, or read release notes — not part of the upstream thread.