vllm - 💡(How to fix) Fix [ROCm]: Docker/CMake build support for gfx1103 (Radeon 780M / RDNA3 APU)

🚀 The feature, motivation and pitch

vLLM's ROCm Docker images (docker/Dockerfile.rocm_base) and the CMake build (CMakeLists.txt → HIP_SUPPORTED_ARCHS) do not currently support gfx1103 (Radeon 780M and related RDNA3 APUs). Users on these integrated GPUs hit two distinct problems when trying to build/run vLLM in a container:

1. CMake rejects the arch. gfx1103 is not listed in HIP_SUPPORTED_ARCHS, so the build fails early even when PYTORCH_ROCM_ARCH=gfx1103 is set.
2. rocBLAS TensileLibrary.dat does not ship gfx1103 kernels in current ROCm releases — this is an upstream AMD limitation rather than a vLLM bug, but it affects the end-to-end experience and is worth documenting.

A proposed patch for (1) — adding gfx1103 to HIP_SUPPORTED_ARCHS plus the corresponding Docker build wiring — was drafted in #31062 (closing in favor of this tracking issue, since I can't validate gfx1103 end-to-end on my current hardware). The diff there can serve as a starting point for anyone with gfx1103 access who wants to pick this up.

Prior testing and reproduction details were contributed by @yurikhan in the #31062 thread.

Alternatives

• Document gfx1103 as explicitly unsupported until rocBLAS ships kernels for it upstream.
• Provide a build path that allows experimental/unsupported archs through with a clear warning.

Additional context

• Affected files in the proposed patch: CMakeLists.txt, docker/Dockerfile.rocm_base.
• Hardware: Radeon 780M / RDNA3 APUs (gfx1103).