search

NVIDIA-Merlin / HugeCTR

HugeCTR is a high efficiency GPU framework designed for Click-Through-Rate (CTR) estimating training
Apache License 2.0
937 stars 200 forks source link

[Question] nv_gpu_cache compiling problem #428

Closed RobertLou closed 11 months ago

RobertLou commented 11 months ago

I'm trying to test nv_gpu_cache with cache_op_sol_test, but encounter strange compiling problems: 1.Changes made in nv_gpu_cache.cu don't seem to affect the results. Even if I add code like printf() or std::cout, the console output doesn't include my new output. Additionally, when I delete functions like _initcache(), it appears that they are still being executed when profiling with nvprof.

  1. When I intentionally introduce syntax errors in nv_gpu_cache.cu, the compiler does show the errors, indicating that it is compiling the file. My compile code is shown below: cmake -DCMAKE_BUILD_TYPE=Release -DSM=75 -DENABLE_INFERENCE=ON .. make -j cache_op_sol_test My running commands is shown below： ./bin/cache_op_sol_test 4096 512 128 128 2 1 0 Environment: Tesla T4 with GPU docker nvcr.io/nvidia/merlin/merlin-hugectr:23.08
yingcanw commented 11 months ago

@RobertLou If you are using ngc docker image, then we have pre-installed all libraries of V23.08 to the default path(/usr/local/hugectr), and add to LD_LIBRARY_PATH. Therefore, if you need to do customized development, please make installthe new library or delete the default installation library and re-specify your library path.