Pull requests / #2283
#2283 setup: opt-in gfx1010 (RX 5700 XT) plus a community bench
open · @Gedankenn · 0 comentarios · En GitHub
BenchmarksSetup & installAMD / HIPNVIDIA / CUDAModels & quantsDocumentationLinux
Descripción
## Summary - Opt-in `STRATA_EXPERIMENTAL_GFX1010=1` so setup can install HIP on an RX 5700 XT (gfx1010). Unset, the default AMD list is unchanged. - Community bench of Coder IQ1_M on that card with engine 0.1.43.1 (`bench/results/2026-10-13-community-rx-5700-xt-coder-iq1_m/`). The HIP arch gate already accepts gfx1010 as unvalidated (from #996 / #1142). Setup still refused the card. This is the gfx1103-shaped hole. ## What it leaves alone CUDA, the default `AMD_ARCHS` list, ready-made engines, and every other HIP architecture. ## How to run it ``` STRATA_EXPERIMENTAL_GFX1010=1 ./setup.sh --yes --family coder --backend hip --gpu 0 --no-start --vision no ``` If the PC also has an AMD iGPU, the start config sets `ROCR_VISIBLE_DEVICES=0`. ROCm is TheRock `gfx101X-dgpu`. ## Bench (this machine) RX 5700 XT 8 GB (drives the desktop) + Renoir iGPU hidden, Ryzen 5 4600G, 46 GB RAM, Linux, HIP source, TheRock 7.10.0a20251120. Coder IQ1_M, 32K, `--spec 4`, int8 KV. | | median | |---|---| | chat decode | 18.0 tok/s (16.7–22.7, n=9) | | 4K prompt | 163.7 tok/s (161.5–163.8) | | 16K prompt | 162.1 tok/s (160.7–163.2) | | agent decode | 17.4 tok/s | Not measured: needle, ctest on the card, 30K prompt. No hipBLASLt table for gfx1010. A different 5700 XT (7600X, ROCm 7.14) is already in docs/OLDER_GPUS.md at 115 / 144 / 149 prompt tok/s (4K / 16K / 30K). This is a second machine, not a rerun. ## Test plan - [x] `python -m unittest tools.test_setup_amd` (39 tests, including `Gfx1010OptIn`) - [x] Model ran on the card; `./bench.sh` wrote the report folder - [ ] Maintainers: default path without the env stays as today
En el sitio
Enlaces a install, modelos, releases.