In plain English
Living notes mention a ROCm 10.0 fine-tune / training GPU-reset class on the R9700 and a first mitigation of `TORCH_BLAS_PREFER_HIPBLASLT=0`. We do not yet have a first-party AMD known-issue URL copied here. Do not treat this as a measured inference fix.
Impact on consumers
Training jobs die with a GPU reset. Inference daily-driver on this site is llama.cpp Vulkan and is a different stack.
Workaround
TODO: confirm the KI URL on the ROCm 10.0.0 notes page, then retry the training step with `TORCH_BLAS_PREFER_HIPBLASLT=0`. Do not apply this to inference because a blog said so.