← 返回事件
持续讨论AI发版

trunk/d2e122ddd079e2386d88ca97681fd1d5a9c28060: Parametrize FP4 scaled_mm tests by recipe instead of platform. (#197740)

图:PyTorch Releases

发生了什么

test_blockwise_nvfp4_compile picked its recipe off torch.version.hip, so on ROCm it ran MXFP4 under an NVFP4 name; ScaledBlas.cpp rejects NVFP4 on ROCm, so the test never could have covered it. The same ternary in test_blockwise_mxfp8_nvfp4_error_messages made mxfp4 unselectable on CUDA, leaving its "recipe == mxfp4 and SM120OrLater" skip permanently dead and MXFP4 untested there. Both tests now parametrize the full…

摘要按规则整理自下方来源原文

为什么在扩散

来源