

November 2025 monthly summary for ROCm/aiter: Stabilized FP8 support by fixing the Torch FP8 global variable initialization path. The bug caused hipGetDeviceProperties to run during static initialization, which could trigger HIP runtime initialization using uninitialized global state and fail in forked child processes. The fix defers hipGetDeviceProperties until it is actually needed, ensuring correct GPU context initialization in both parent and forked children. This reduces risk of hipErrorInvalidValue in FP8 workflows and improves overall reliability for FP8 training/inference.
November 2025 monthly summary for ROCm/aiter: Stabilized FP8 support by fixing the Torch FP8 global variable initialization path. The bug caused hipGetDeviceProperties to run during static initialization, which could trigger HIP runtime initialization using uninitialized global state and fail in forked child processes. The fix defers hipGetDeviceProperties until it is actually needed, ensuring correct GPU context initialization in both parent and forked children. This reduces risk of hipErrorInvalidValue in FP8 workflows and improves overall reliability for FP8 training/inference.
Overview of all repositories you've contributed to across your timeline