Pull requests / #3272

#3272 [Other] opt-in gfx1010 in setup (RX 5700 XT) plus a community bench

open · @Gedankenn · 0 Kommentare · Auf GitHub

BenchmarksSetup & installAMD / HIPNVIDIA / CUDAModels & quantsWindowsLinux

Beschreibung

## Problem and result

Setup refused an RX 5700 XT (gfx1010) even though the HIP arch gate already compiles that card as unvalidated. With `STRATA_EXPERIMENTAL_GFX1010=1`, setup treats gfx1010 like the gfx1103 opt-in: TheRock `gfx101X-dgpu` wheels, the card is usable, default AMD list unchanged. Unset, setup still refuses it.

## Related discussion

Continues https://github.com/Niko1221/Strata/pull/2283 (closed in the 0.2.0-beta tracker reset, not rejected; `/reopen` passed the check but GitHub would not reopen that number). Earlier notes: HIP already accepts gfx1010 as unvalidated (#996 / #1142). Review on #2283: `test_setup_rocm_pins` KeyError on `gfx101X-dgpu`, and `ROCR_VISIBLE_DEVICES=0` hardcoded.

## Approach

Same opt-in shape as gfx1103 / gfx1150: index in `ROCM_INDEXES` always, arch in `AMD_ARCHS` only when the env is set. `ROCR_VISIBLE_DEVICES` is now the chosen card's HIP index, and only when another unsupported AMD device is listed (typically an iGPU). It is in `SETUP_ENV` so a re-run without the sibling drops it. CUDA, the default `AMD_ARCHS` list, ready-made engines, and every other HIP architecture are unchanged.

Community bench folder from this machine is included (engine 0.1.43.1; not a rerun of the 7600X numbers already in OLDER_GPUS.md).

## Validation

**Passed:**
- `python -m unittest tools.test_setup_amd tools.test_setup_rocm_pins tools.test_setup_start_gpus_amd` on 266413610431f127f6bb0a07dcaee31a39025d8d (rebased on current main). 67 tests, 0 failures. Includes `Gfx1010OptIn` (refuse without opt-in, index present, opt-in admits the card, ROCR index vs iGPU / lone card / supported sibling). `test_setup_rocm_pins` skips `gfx101X-dgpu` the same way it skips gfx1150 (no saved listing).
- Hardware (older revision, not this SHA): RX 5700 XT 8 GB (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, engine 0.1.43.1. Chat decode 18.0 tok/s (16.7–22.7, n=9); 4K prompt 163.7 tok/s; 16K prompt 162.1 tok/s; agent decode 17.4 tok/s. Report: `bench/results/2026-10-13-community-rx-5700-xt-coder-iq1_m/`.

**Failed:**
none observed on the unit tests above.

**Not run:**
- HIP rebuild / `./bench.sh` / needle / 30K prompt / ctest on the card against this SHA (0.2.0-beta tree). The numbers above are 0.1.43.1. Niko already noted the engine target compiles for gfx1010 on the then-current tree.
- CUDA / SYCL / Windows HIP.

## Limits

Opt-in, unsupported. No hipBLASLt table for gfx1010. `ROCR_VISIBLE_DEVICES` is set only when an unsupported AMD sibling is present; a lone 5700 XT is left alone. Default path without the env is unchanged.

Mehr auf der Site

Links zu Install, Modellen, Releases.