Cudnn error? I don't know why

[muxifn]

Attempting to switch between multiple versions of cudnn will result in the following error, it is unclear what caused it. Can you help me?

[libprotobuf ERROR external/com_google_protobuf/src/google/protobuf/text_format.cc:337] Error parsing text-format xla.gpu.DeviceHloInstructionProfiles: 1463:15: Message type "xla.gpu.HloInstructionProfile" has no field named "entries".
2024-01-14 03:06:41.613349: F external/xla/xla/service/gpu/gpu_hlo_cost_analysis.cc:289] Check failed: tsl::protobuf::TextFormat::ParseFromString( std::string(kDeviceHloOpProfiles), &all_device_profiles)

[inter1965]

that's beacause in "xla/service/gpu/hlo_op_profiles.h" there's a huge embedded string kDeviceHloOpProfiles, it's fine when you complies it with gcc or clang, however with msvc you would trigger "C2026" error (msvc error c2026: string too big), you need to use something like bin2c to embedded it or change the toolchain to clang-cl.

[weirenorweiren]

Hi @muxifn, I encountered the same error when trying to set up https://whls.blob.core.windows.net/unstable/cuda121/jaxlib-0.4.14+cuda12.cudnn89-cp311-cp311-win_amd64.whl in #35 (comment). Did you manage to solve it?

Hi @inter1965, could you elaborate more on how to proceed? I have no prior knowledge on the terms you mentioned and I just know they might be relevant to the installation of Visual Studio 2022 (which is required by CUDA 12.1 according to https://docs.nvidia.com/deeplearning/cudnn/latest/reference/support-matrix.html#).