Radeon VII users seek GPT-OSS compatibility on legacy ROCm

The Radeon VII remains a legendary card for local inference, but its "legacy" status in the ROCm ecosystem is creating a software bottleneck that prevents users from accessing the latest open-weights breakthroughs.

• –Setting HSA_OVERRIDE_GFX_VERSION=9.0.6 is the critical "secret handshake" for modern llama.cpp builds to recognize the Vega 20 architecture.
• –"Unknown architecture" errors typically stem from a binary that lacks support for newer GGUF metadata keys introduced by recent model releases like GPT-OSS.
• –The 16GB VRAM limit makes GPT-OSS-20B the primary target for this user, though its sparse MoE architecture requires up-to-date kernel support that older binaries lack.
• –Users are better off tracking the latest llama.cpp source and using custom compiler flags (AMDGPU_TARGETS=gfx906) rather than downgrading to "safe" older versions that are blind to new model types.