In the last few days, I was experimenting with ZenithO-o/fursuit-detection, after sorting some photos from a fursuit walk. I couldn’t get it to work/run, no matter how hard I tried, with the system packages. (Debian stable is somewhat old, granted, which leads to some problems).

So since venv wouldn’t do it’s job, I thought about (ana|mini)conda again. Only to find out now there’s also miniforge. And apparently a project with a faster dependency resolver, mamba, split up. Its dependency resolver already re-integrated into conda (?), and then there’s micromamba which is a standalone executable compiled from C++ (?)… I already wasted some hours researching that rabbit hole. And apparently you can’t use that stuff without putting some stuff into your .bashrc, “activate only when needed” doesn’t seem to be a supported usecase? (I didn’t want to put even more time into this, but yes, looking at what’s inside `.bashrc’, I could simply do this manually…).

So anyways. Next step was searching for the required packages in conda-forge. I found some very outdated guides on the internet, which installed some things manually. I simply went with micromamba install tensorflow-gpu - and hey, it works! Or so I thought…

Running the run_on_images script gave me a

tensorflow.python.framework.errors_impl.InternalError: Graph execution error:

Detected at node MultiscaleGridAnchorGenerator/GridAnchorGenerator/mul_3 defined at (most recent call last):
<stack traces unavailable>
Detected at node MultiscaleGridAnchorGenerator/GridAnchorGenerator/mul_3 defined at (most recent call last):
<stack traces unavailable>
2 root error(s) found.
  (0) INTERNAL:  'cuLaunchKernel(function, gridX, gridY, gridZ, blockX, blockY, blockZ, 0, reinterpret_cast<CUstream>(stream), params, nullptr)' failed with 'CUDA_ERROR_INVALID_HANDLE'
	 [[ { {node MultiscaleGridAnchorGenerator/GridAnchorGenerator/mul_3 } } ]]
	 [[StatefulPartitionedCall/map/while/loop_body_control/_430/_23]]
  (1) INTERNAL:  'cuLaunchKernel(function, gridX, gridY, gridZ, blockX, blockY, blockZ, 0, reinterpret_cast<CUstream>(stream), params, nullptr)' failed with 'CUDA_ERROR_INVALID_HANDLE'
	 [[ { {node MultiscaleGridAnchorGenerator/GridAnchorGenerator/mul_3 } } ]]
0 successful operations.
0 derived errors ignored. [Op:__inference_restored_function_body_41075]

A very useless error message. I did a lot of fruitless internet searches. I then noticed the author of the script limited logging. So I removed that line. With that, I suddenly got a more promising

2024-10-03 16:04:42.816559: W tensorflow/compiler/mlir/tools/kernel_gen/tf_gpu_runtime_wrappers.cc:40] 'cuModuleLoadData(&module, data)' failed with 'CUDA_ERROR_NO_BINARY_FOR_GPU'
2024-10-03 16:04:42.816603: W tensorflow/compiler/mlir/tools/kernel_gen/tf_gpu_runtime_wrappers.cc:40] 'cuModuleGetFunction(&function, module, kernel_name)' failed with 'CUDA_ERROR_INVALID_HANDLE'

Which didn’t help me much either itself. But then I spotted this at the beginning of the script

2024-10-03 16:04:36.034469: W tensorflow/core/common_runtime/gpu/gpu_device.cc:2432] TensorFlow was not built with CUDA kernel binaries compatible with compute capability 5.2. CUDA kernels will be jit-compiled from PTX, which could take 30 minutes or longer.

So… my GPU was too old. And apparently something went wrong with the JIT compilation. So I step-by-step installed older tensorflow-gpu packages until I arrived at tensorflow-gpu~=2.14.0. Mind you, this whole process took half a day. And even then, I wasn’t spared:

2024-10-03 18:16:03.567311: W tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc:559] libdevice is required by this HLO module but was not found at ./libdevice.10.bc
error: libdevice not found at ./libdevice.10.bc
2024-10-03 18:16:03.568777: W tensorflow/core/framework/op_kernel.cc:1827] UNKNOWN: JIT compilation failed.
2024-10-03 18:16:03.568846: W tensorflow/core/framework/op_kernel.cc:1827] UNKNOWN: JIT compilation failed.

Luckily, for that one, I found a solition pretty quickly: You have to copy a file to a subdirectory in the execution directory:

cp ${PUT_ENV_PATH_HERE}/lib/libdevice.10.bc ./cuda_sdk_lib/nvvm/libdevice/

And then, FINALLY!!!, this sh** works. Half a day wasted, and a lot of angry shouts were emitted.