HSA_OVERRIDE for RDNA2 Cousins

Only Navi 21 cards report as gfx1030. The smaller RDNA2 dies use different LLVM targets:

Many ROCm libraries only ship precompiled kernels for a subset of targets. Because every RDNA2 card shares the same instruction set family, you can tell ROCm to treat your card as gfx1030 and reuse the gfx1030 kernels.

The override

export HSA_OVERRIDE_GFX_VERSION=10.3.0

10.3.0 maps to gfx1030. Set it in the shell (or systemd unit / container env) before launching any HIP/ROCm program:

HSA_OVERRIDE_GFX_VERSION=10.3.0 python my_inference_script.py

To make it permanent for your user:

echo 'export HSA_OVERRIDE_GFX_VERSION=10.3.0' >> ~/.bashrc

Why this works (and its limits)

• RDNA2 GPUs (gfx1030–gfx1036) are binary-compatible enough that gfx1030 kernels execute correctly on the smaller dies for the vast majority of ML ops.
• It is still a workaround. AMD does not officially validate it, and you may hit edge cases — particularly in hand-tuned assembly kernels or libraries that query the exact arch at runtime.
• Performance-tuned kernels (e.g. in rocBLAS/Tensile) were tuned for Navi 21’s CU count and cache; on a smaller die they run correctly but may be sub-optimal.

Building instead of overriding

If you compile a library yourself, prefer building for your real target so you get correctly-tuned kernels, e.g.:

# Build for multiple RDNA2 targets at once
export PYTORCH_ROCM_ARCH="gfx1030;gfx1031;gfx1032"
# or for a single card
export AMDGPU_TARGETS=gfx1032

For llama.cpp:

cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release

Verifying it took effect

HSA_OVERRIDE_GFX_VERSION=10.3.0 python -c \
  "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

If PyTorch previously errored with "no kernel image is available for execution on the device" and now prints True, the override is working.