[Bug] report missing libtvm.so when specify the TVM_LIBRARY_PATH on gpu

[vfdff]

• It reports missing libtvm_allvisible.so/libtvm.so when I compile the tvm base on the source according https://tvm.hyper.ai/docs/install/from_source/

[100%] Linking CXX shared library libtvm_runtime.so
[100%] Built target tvm_objs
make[2]: *** No rule to make target '/usr/lib/llvm-14/lib/libPolly.a', needed by 'libtvm_allvisible.so'.  Stop.
make[2]: *** No rule to make target '/usr/lib/llvm-14/lib/libPolly.a', needed by 'libtvm.so'.  Stop.
make[1]: *** [CMakeFiles/Makefile2:260: CMakeFiles/tvm_allvisible.dir/all] Error 2

• I try both the config set(USE_CUDA OFF) and set(USE_CUDA ON), they have the same errors

Expected behavior

compile successfully and get libtvm.so and libtvm_runtime.so

Actual behavior

• compile failed
  

• the libtvm.so and libtvm_runtime.so exist under $TVM_LIBRARY_PATH
  

Environment （83, gpu machine）

Linux version 5.15.0-46-generic (buildd@lcy02-amd64-007) (gcc (Ubuntu 9.4.0-1ubuntu120.04.1) 9.4.0, GNU ld (GNU Binutils for Ubuntu) 2.34) #4920.04.1-Ubuntu SMP Thu Aug 4 19:15:44 UTC 2022 （cat /proc/version）

Steps to reproduce

accord https://tvm.hyper.ai/docs/install/from_source/

git clone --recursive https://github.com/apache/tvm tvm --depth=1
git fetch --update-shallow
mkdir build
cp cmake/config.cmake build
cmake ..
make -j

[vfdff]

oh, I rebuild it after clean the value of TVM_LIBRARY_PATH (export TVM_LIBRARY_PATH=)， it still only generates libtvm_runtime.so under the build Folders.

[lipeng821020589]

@vfdff hi, recommended that you use LLVM 15. The problem may be a known problem(https://discourse.llvm.org/t/ubuntu-packages-with-llvm-14-then-15-suddenly-looking-for-libpolly-a/67655).

[tqchen]

seems we now need to apt-get install libpolly-dev along with llvm dev here

[vfdff]

Thanks all, and I find the llvm-14 can also match the tvm, refer to #17558

(base) root@j00595921debug2-cc95c9977-q752v:/home/zhongyunde/tvm_codes/apache-tvm-src-v0.18.0/build# ldd libtvm.so
	linux-vdso.so.1 (0x00007ffe29a9c000)
	libnvrtc.so.12 => /usr/local/cuda-12.2/lib64/libnvrtc.so.12 (0x00007ffb5c800000)
	libLLVM-14.so.1 => /usr/lib/llvm-14/lib/libLLVM-14.so.1 (0x00007ffb55f2e000)

• On gpu should install nvcc first to enable USE_CUDA ON, then install llama-cpp-python

CUDACXX=/usr/local/cuda/bin/nvcc CMAKE_ARGS="-DGGML_CUDA=on -DLLAVA_BUILD=off" FORCE_CMAKE=1 pip install llama-cpp-python --no-cache-dir --force-reinstall --upgrad

[vfdff]

Base on tvm v0.21, just comment the setting of USE_LLVM looks also work fine.

   # set(USE_LLVM  "/usr/bin/llvm-config-14 --link-static")

[vfdff]

Base on old tvm v0.7, should set the detail path

set(USE_CUDA  "/usr/local/cuda")

[vfdff]

• another works around: apt-get install libclang-common-12-dev as we may failed with apt-get install libpolly-12-dev
  refer to https://bugs.launchpad.net/ubuntu/+source/llvm-toolchain-11/+bug/1934927