Pull requests / #1313
#1313 hipBLASLt table for gfx1201 at 100200 (R9700 calibrated, +15% prefill)
closed · @Faks · 0 comentarios · En GitHub
Setup & installAMD / HIPModels & quants
Descripción
## Why Upstream ships gfx1201 tables for 100202/100500 but nothing for 100200 — exactly what TheRock 7.10-era venv wheels (what setup installs, what the engine loads) report. Without a table the engine runs plain hipBLAS on every dense GEMM. ## What - `tools/hip/gfx1201-hipblaslt-100200.txt`: 26 rows via `tune_hipblaslt` (cases from the gfx1100-100200 template), calibrated on Radeon AI PRO R9700. - `tools/hip/gfx1201-hipblaslt-100200.PROVENANCE.md`: exact command, the venv-vs-system lib trap, verify output, A/B numbers. ## Proof - Parity: `hip_prefill_hipblaslt_gemm` with the table — PASS, `fallbacks=0`. - Live A/B (Qwen3.8-Flash-Next IQ3_S, int8+WMMA, fresh 52.9K prompt, R9700/5950X): prefill 2294.4 vs 2000.9 plain (+15%), decode neutral. - Failure recorded too: first calibration ran against system libs (stamped 100401, engine refused it). Lesson in PROVENANCE.md. ## Risk Near zero: additive files only, the engine refuses mismatched tables by itself. Worst case is status quo.
En el sitio
Enlaces a install, modelos, releases.