Pull requests / #1127

#1127 hip: rebased support for Radeon gfx1151 APUs

open · @DTU-travelpals · 0 评论 · 在 GitHub 查看

Setup & installAMD / HIPModels & quantsWindows

描述

## Summary

Focused rebase of the remaining changes from #895 onto the repository's
rewritten `main`.

Most gfx1151 support is already present in current `main`, including Strix Halo
detection, unified-memory accounting, compilation support, Windows packaging,
setup integration, kernel compatibility, and documentation.

This PR retains the two pieces that were still missing:

- preserve Q2_0 signed zero on gfx1151 HIP compilers when a negative scale is
  multiplied by code zero;
- add the Radeon 8060S hipBLASLt 1.5.0 (`100500`) tuning table calibrated with
  AMD's 2026-10-04 nightly SDK.

It also corrects the setup/install documentation to name experimental Strix
Halo support consistently.

## Validation

The original changes were validated on:

- Ryzen AI Max+ 395;
- Radeon 8060S (`gfx1151`);
- 128 GB unified memory;
- 116.0 GiB GPU-addressable GTT pool;
- HIP runtime 7.17.26392;
- hipBLASLt 1.5.0.

The tuning table's smoke test exercised BF16 and F16 rows with outputs matching
hipBLASEx and `fallbacks=0`. End-to-end prompt performance with this specific
1.5.0 table was not measured.

After rebasing:

- `python tools/test_setup_amd.py`: 30 tests passed;
- `python tools/test_setup_strix_halo.py`: 34 tests passed;
- `git diff --check`: clean.

## AI assistance disclosure

Implementation review, documentation preparation, and commit organization were
assisted by AI. Hardware builds, GPU tests, and reported measurements were run
on the Radeon 8060S system described above.

站内延伸阅读

链到安装、模型与版本说明,便于 SEO/GEO,非官方 issue 正文。