Repository navigation

Add AMD ROCm/HIP backend for SmolVLA and π0.5 on gfx1151 - #35

Merged

khanhnd61-vr merged 9 commits into

VinRobotics:mainfrom

singlecatlmx:pr/rocm-hip-v04

Oct 28, 2026

Merged

Add AMD ROCm/HIP backend for SmolVLA and π0.5 on gfx1151#35
khanhnd61-vr merged 9 commits into
VinRobotics:mainfrom
singlecatlmx:pr/rocm-hip-v04

Commits on Oct 3, 2026
- Add ROCm/HIP backend selection for VLA models
  singlecatlmx
  committed
  b1f0fa8
- Document validated gfx1151 ROCm/HIP coverage
  singlecatlmx
  committed
  9803295
- Use GGML_USE_HIP for VLA HIP backend guards
  singlecatlmx
  committed
  30fb537
Commits on Oct 7, 2026
- Merge upstream main into ROCm/HIP PR branch
  singlecatlmx
  committed
  bb94652
- Refresh ROCm validation after upstream merge
  singlecatlmx
  committed
  002a41c
Commits on Oct 27, 2026
- Merge origin/main into pr/rocm-hip-v04
  khanhnd61-vr
  committed
  6137589
- fix(act): keep the CUDA runtime query out of HIP builds
  khanhnd61-vr
  committed
  6d54233
Commits on Oct 28, 2026
- perf(smolvla): widen LM GEMM weights and cache F16 cross K/V on ROCm/HIP as on CUDA
  khanhnd61-vr
  committed
  d0e5d00
- docs(rocm): validate and benchmark every model on the Radeon RX 9060 XT (gfx1200)
  khanhnd61-vr
  committed
  00249cf