Pull requests / #2147
#2147 setup: opt-in gfx1150 (Radeon 890M / 880M, Strix Point) as a unified-memory AMD card
open · @onionbutties · 0 Kommentare · Auf GitHub
BenchmarksSetup & installAMD / HIPWindows
Beschreibung
Issue: #1217 (the gfx1150 build/run report) - setup does not install for gfx1150 yet
## Summary
An opt-in, `STRATA_EXPERIMENTAL_GFX1150=1`, that lets setup install for the Radeon 890M / 880M iGPU (Ryzen AI 300,
Strix Point, gfx1150) as a unified-memory AMD card like Strix Halo. The engine already builds and runs for gfx1150
(0.1.40.2, #1217); only the installer did not know it.
## What changed
- `setup.py`
- `GFX1150_OPT_IN` and gfx1150 in `AMD_ARCHS` / `ROCM_INDEXES` (TheRock `gfx1150` index), gated so the default path
is byte-identical when the variable is unset.
- `is_uma_apu(g)`: the shared unified-memory handling now covers Strix Halo (gfx1151), gfx1150 and the gfx1103
opt-in (was Strix Halo plus gfx1103 inline). `amd_apply_uma`, `igpu_notes` and `amd_gpus_win` use it.
- `gfx1150_notes`: the unified-memory notes for this chip (never the Strix Halo text).
- The install writes the chip's exact speed switches into the config's `env`, beside `STRATA_HIPBLASLT_TUNING`
(`GFX1150_SWITCHES`): the gfx1151 arch-default set (`src/core/arch_defaults.cpp`) less `STRATA_HCD_EXACT`, which
copies the gfx1151 table's solutions 1176 / 1177 and does not fit a gfx1150 table. The Radeon 890M report measured
the same greedy answers with all 17 on and with all 17 off, decode +2% and prompts +3-6%. The set is in
`SETUP_ENV`, so a re-run refreshes it and it does not linger when the opt-in is turned off.
- `tools/test_setup_strix_halo.py`: `Gfx1150Switches` keeps `GFX1150_SWITCHES` equal to the gfx1151 table minus
`STRATA_HCD_EXACT`, asserts the engine's own default path for gfx1150 stays empty
(`arch_default_env("gfx1150").empty()`), and checks the `SETUP_ENV` wiring.
## Extra Notes
- No release ships a ready-made Windows engine for gfx1150, so the note tells the user to build one with
`STRATA_HIP_ARCHS=gfx1150 tools\hip\build_windows.bat` and start setup with `--prebuilt dist\`.
- Default path is byte-identical when the variable is unset; `tools/test_setup_strix_halo.py` (40),
`tools/test_setup_amd.py` (36) and `tools/test_setup_choices.py` (47) pass.
- Verified end to end on a Radeon 890M (Ryzen AI 9 HX 370, 96 GB, Windows 11): the engine builds for gfx1150, setup
accepts the card, and the model runs with all 24,576 experts in the GPU pool at about 11 tokens/s decode.
- The engine's arch defaults are deliberately left alone (gfx1150 stays out of `arch_defaults.cpp`); the switches are
written by setup instead.
- The gfx1150 hipBLASLt tuning table is #2148.
Mehr auf der Site
Links zu Install, Modellen, Releases.