#6496 [Other] Windows HIP rdna2 package (0.1.63) works on a real RX 6900 XT (gfx1030)

open · @ByteJuggler · 0 comentários · No GitHub

BenchmarksSetup & installAMD / HIPModels & quantsDocumentationWindowsLinux

Descrição

## What was tested

`docs/AMD_HIP.md` says the Windows `rdna2` package is built but not tested on a real RX 6000 card, and asks for a
report. This is one.

- **Package:** `strata-engine-0.1.63-windows-x86_64-hip-rdna2.tar.gz` from the v0.2.17 release. Its SHA-256 matches
  the release's `SHA256SUMS.txt`.
- **Card:** AMD Radeon RX 6900 XT 16 GB (gfx1030, 40 CUs), AMD Software 26.11.2 (driver 32.0.21045.29003).
- **PC:** Intel i7-5930K (AVX2), 192 GB DDR4-2400, Windows 11 Pro 23H2 (build 22631.6199).
- **Model:** Qwen3.8-Flash-Next IQ3_S with its MTP draft layer, `--spec 4 --spec-min-p 0.5`, with
  `STRATA_HIP_PROMPT_F16=1` (the engine's own tip on gfx103x) and `STRATA_ARENA_PIN_GIB=auto`.

## Results

- `strata-device.exe --list-devices`: exit 0, `device 0: AMD Radeon RX 6900 XT, arch gfx1030, 16.0 GiB, wave32`.
- `strata-device.exe --selftest`: exit 0, `strata-device selftest OK`. It reports being compiled for
  gfx1030, gfx1031 and gfx1032 and uses the package's own `amdhip64_7.dll`. The memory plan fits, and an
  over-allocation is refused as it should be.
- **Engine start and decode:** 16 requests over 2 rounds (3 short chats and a 6,345-token prompt, greedy), all
  answered, exit 0. The engine log shows the package's own HIP runtime and `mtp: draft layer loaded, 835 MiB of VRAM`.
  - First answer: 34-token prompt, 256 tokens generated at 34.6 tok/s, 179 of 212 drafts accepted.
  - Long prompt: 6,345 tokens read in 6.7 s (942 tok/s), then 132 tokens at 31.7 tok/s.
  - Decode medians without / with MTP: 26.6 / 33.8, 30.6 / 35.5, 29.8 / 33.6 tok/s (chats), 28.4 / 31.3 tok/s (long
    prompt). 62-63% of drafts accepted.
- **Workarounds:** none, beyond the one step the docs already give: the package's `rocm\bin` must be on PATH to run
  `strata-device.exe` by hand. The engine itself found its runtime without it.
- The same card ran this package as the "before" in #6433. There the only refusal was the one that PR fixes, an MTP
  head with 256 experts instead of 512.

## Not tested

- gfx1031 and gfx1032 cards (RX 6700 XT, RX 6600 series).
- The Linux `rx6000` package.
- Vision (the picture readers in the package).
- The app installer; the package was unpacked and run by hand.

## Suggestion

The RDNA2 row and section in `docs/AMD_HIP.md` could now say the Windows `rdna2` package was tested on a gfx1030 card
(RX 6900 XT), with the Linux package and gfx1031/gfx1032 still untested. One question: the docs list the `rdna2`
package as gfx1030/gfx1031, while its selftest says it is compiled for gfx1032 too. Is the gfx1032 code meant to be
there, and should the docs mention it?

No site

Links install, modelos, releases.