Pull requests / #917

#917 bench: community report, Qwen3.8-Flash-Next-GSQ-RCO Q2_0 (Strix Halo/Radeon 8060S gfx1151 + 128 GB VRAM)

closed · @DTU-travelpals · 0 commentaires · Sur GitHub

BenchmarksAMD / HIPModels & quants

Description

This PR is dependent on PR #895, in addition it covers the hipBLASLt 1.5 table from PR #755 adjusted for gfx1151 (big performance boost for prefill).
Test cases cover in addition the merged PR #826 and PR #820.

Record the interactive Q2_0 and UD-IQ4_XS observations plus the matched A-E Q2_0 prefill/decode comparison. Include the generated configs, build provenance, raw request JSON and engine logs.

Sur le site

Liens install, modèles, releases.