Repository navigation

AMD MI300X (CDNA3) support — Pi0.5 at 35.1 ms BF16, GROOT N1.7 at 19.6 ms - #214

Merged

LiangSu8899 merged 13 commits into

flashrt-project:mainfrom

yiannisha:feat/amd-mi300x

Sep 30, 2026

Merged

AMD MI300X (CDNA3) support — Pi0.5 at 35.1 ms BF16, GROOT N1.7 at 19.6 ms#214
LiangSu8899 merged 13 commits into
flashrt-project:mainfrom
yiannisha:feat/amd-mi300x

Commits on Sep 24, 2026
- feat(amd): compile the HIP backend for gfx942
  yiannisha
  committed
  07b069e
- feat(amd): route CDNA3 through shared capabilities
  yiannisha
  committed
  be57d60
- fix(amd): match the ROCm 7 AITER attention ABI
  yiannisha
  committed
  6160f63
- feat(amd): run GROOT N1.7 on CDNA3
  yiannisha
  committed
  a363df9
- perf(amd): tune the Pi0.5 decoder for gfx942
  yiannisha
  committed
  3ade784
- test(amd): cover CDNA3 capabilities and routing
  yiannisha
  committed
  97ee61b
- test(amd): validate gfx942 kernel parity
  yiannisha
  committed
  7dbad91
- test(amd): exercise MI300X model paths
  yiannisha
  committed
  5af6ecb
- bench(amd): add real-input MI300X latency drivers
  yiannisha
  committed
  b079f29
- docs(amd): document gfx942 platform support
  yiannisha
  committed
  4f2fdee
- docs(amd): report MI300X model results
  yiannisha
  committed
  cc037b9
Commits on Sep 25, 2026
- Merge remote-tracking branch 'upstream/main' into feat/amd-mi300x
  yiannisha
  committed
  4094166
Commits on Sep 29, 2026
- Fix AMD packed BF16 routing with shape and layout capabilities
  yiannisha
  committed
  df6dcc9