Why require native instead of HSA_OVERRIDE?

Override gave FA-off, top_k CPU fallback tanking decode, and random mid-inference core dumps. For 35B split at 70K+ that is unusable. Native was slower to achieve but stable, with rocminfo proving no spoofing.

Is head_dim 512 a universal fix I can copy?

No. Conditional on amd_mfma_available and this HIP + commit. Validate failure mode first. Method generalizes, number doesn't.

Why did SGLang and vLLM start working too?

Unknown. Flagged honestly. Likely parallel ROCm-level changes untracked. If you understand the mechanics, author invites explanation.

From Override to Native: Robust ROCm gfx1031 Without HSA_OVERRIDE.

Home Profile Experiments Contact Experiment 001 Experiment 002 Experiment 003 Experiment 004 Experiment 005 Experiment 006 Experiment 007 Experiment 008 Experiment 009 Experiment 010 Experiment 011 Experiment 012 Experiment 013 Experiment 014 Experiment 015 Experiment 016 Experiment 017